CN-121997455-A - Spacecraft repeatable service life prediction method based on learning dynamic graph neural network

CN121997455ACN 121997455 ACN121997455 ACN 121997455ACN-121997455-A

Abstract

A spacecraft repeatable service life prediction method based on a learning dynamic graph neural network is used for training a space-time characteristic prediction network comprising a plurality of space-time blocks by collecting time sequence data of a plurality of sensors and generating a training set in an off-line stage, and carrying out real-time service life prediction by the trained space-time characteristic prediction network in an on-line stage. The method for predicting the repeatable service life of the graph structure can be used for end-to-end learning and dynamically adjusting, solves the problem that the fixed graph structure lacks flexibility, and realizes more accurate service life prediction with stronger generalization capability.

Inventors

HE QINGBO
Qiao Xinhang
Jiao Zehang
LI WENQU
HUANG JINGKAI

Assignees

上海交通大学

Dates

Publication Date: 20260508
Application Date: 20260124

Claims (7)

1. A spacecraft repeatable service life prediction method based on a learning dynamic graph neural network is characterized by comprising the steps of collecting multi-sensor time sequence data and generating a training set in an off-line stage, and training a space-time characteristic prediction network comprising a plurality of space-time blocks; the space-time characteristic prediction network comprises an input layer, K cascaded space-time blocks, a characteristic flattening layer, a full-connection regression network and a service life prediction layer.
2. The method for predicting the repeatable service life of the spacecraft based on the learning dynamic graph neural network is characterized in that the input layer performs standardization processing on sensor data according to time sequence data information synchronously acquired by multiple sensors and segments the sensor data in a sliding window mode to obtain a multi-sensor time sequence input matrix, each cascade space-time block performs space relation modeling and time dependency relation modeling processing according to a multi-sensor time sequence input matrix from the input layer or the previous space-time block and based on a dynamically changed sensor association relation to obtain a space-time feature tensor fusing space features and time features, the feature flattening layer performs flattening processing on the space-time feature tensor information in a time dimension and a sensor dimension according to the space-time feature tensor information to obtain feature vectors for life regression, the full-connection regression network performs feature transformation processing according to the feature vector information to obtain high-order feature representation related to the life degradation state of the spacecraft, and the life prediction layer performs regression prediction processing according to the high-order feature representation information to obtain the repeatable service life prediction value of the spacecraft.
3. The spacecraft repeatable service life prediction method based on the learning dynamic image neural network of claim 1, wherein the space-time block comprises a learning dynamic image convolution sub-module and a time convolution sub-module (TCN), the learning dynamic image convolution sub-module executes image convolution operation according to an initial adjacent matrix, a sample specific dynamic matrix and a global learning matrix of the TCN from an input layer or a previous stage space-time block to aggregate space correlation information to obtain a space feature reflecting a sensor space dependence relation, and the TCN adopts a causal expansion time convolution mode to model degradation characteristics in a time sequence according to the space feature information to obtain a time sequence feature containing the time dependence relation.
4. The spacecraft repeatable service life prediction method based on the learning dynamic graph neural network of claim 3, wherein the initial adjacency matrix is obtained by calculating initial connection strength between sensor nodes according to statistical correlation information of multi-sensor historical operation data; The global learning matrix performs self-adaptive optimization processing on the adjacent relation according to parameter updating information obtained by back propagation in the training process to obtain a global learning matrix for representing the global space dependency relation; The sample specific dynamic matrix is obtained by generating a dynamic connection relation between nodes through embedding mapping and relevance calculation according to an initial adjacent matrix of TCN from an input layer or a previous stage of space-time block.
5. The method for predicting the repeatable usable life of a spacecraft based on a neural network of a leachable dynamic diagram according to claim 3 or 4, wherein the spatiotemporal feature tensor output after further processing by TCN in the kth spatiotemporal block, , Where k is the layer index of the spatio-temporal block, A multi-sensor spatiotemporal feature tensor output for the k-1 th spatiotemporal block, for use as an input feature for the current spatiotemporal block, For the intermediate spatial feature tensor obtained after processing by the learner dynamics graph convolution sub-module in the kth temporal-spatial block, In order to learn the dynamic graph convolution sub-module, Is a time convolution network sub-module; The training uses Mean Square Error (MSE) as a loss function, and optimizes the model ownership weight in an end-to-end manner.
6. The method for predicting the repeatable service life of the spacecraft based on the learning dynamic image neural network of claim 1, wherein the online stage is used for comparing the repeatable service life predicted value of the spacecraft with a preset flight life threshold, and the flight life threshold is determined according to the working condition requirement, the safety margin and the maintenance strategy of the next flight task.
7. The spacecraft repeatable service life prediction method based on the learning dynamic graph neural network according to claim 1, wherein the fully connected regression network predicts and obtains the repeatable service life prediction value according to the high-order fusion characteristic tensor after flattening in time and space dimensions, and specifically comprises the following steps: Wherein: Represents a flattening operation, and the flattening device, And For weight and bias.

Description

Spacecraft repeatable service life prediction method based on learning dynamic graph neural network Technical Field The invention relates to a technology in the field of aerospace, in particular to a spacecraft repeatable service life prediction method based on a learnable dynamic graph neural network. Background In the process of spacecraft health management, sensors are generally arranged on a plurality of key subsystems, and operation data such as temperature, pressure, vibration, electrical parameters and the like are collected in real time. These multisource sensor data record the operational state and degradation process of the spacecraft in time series, and also implies the coupling relationship between the different sensors. The existing life prediction technology based on the graph neural network generally adopts a predefined static graph structure, and cannot adaptively adjust the association relation between sensors according to different working conditions, different repeated use stages and the change of degradation degree of a spacecraft. The lifetime assessment technology based on the coupling of the image field and the physical field does not consider the technical means of dynamically adjusting the association structure between the features due to the change of the operating condition or the degradation stage of the equipment, and the dynamic coupling relation generated by the multi-sensor system along with the state change in the lifetime evolution process cannot be described. Disclosure of Invention Aiming at the defects that the existing life prediction technology depends on a predefined adjacent matrix, cannot be adaptively adjusted according to an input sample or a degradation stage, and is difficult to capture the dynamic change relation among sensors, so that the life prediction precision and generalization capability are limited, the invention provides a spacecraft repeatable service life prediction method based on a learnable dynamic graph neural network, which can learn from end to end and dynamically adjust a graph structure, solves the problem that a fixed graph structure lacks flexibility, and realizes more accurate life prediction with stronger generalization capability. The invention is realized by the following technical scheme: the invention relates to a spacecraft repeatable service life prediction method based on a learnable dynamic graph neural network, which is used for training a space-time characteristic prediction network comprising a plurality of space-time blocks by collecting multi-sensor time sequence data and generating a training set in an off-line stage; and in the online stage, carrying out real-time life prediction through the trained space-time characteristic prediction network. The training set is obtained by carrying out data cleaning and preprocessing on the time sequence data of the multiple sensors. The space-time characteristic prediction network comprises an input layer, a plurality of cascade space-time blocks, a characteristic flattening layer, a full-connection regression network and a life prediction layer, wherein the input layer performs standardization processing on sensor data according to time sequence data information synchronously acquired by a plurality of sensors and segments the sensor data in a sliding window mode to obtain a multi-sensor time sequence input matrix, each cascade space-time block performs space relation modeling and time dependency relation modeling processing according to a multi-sensor time sequence input matrix from the input layer or a previous space-time block based on a dynamically changed sensor association relation to obtain a space-time characteristic tensor integrating space characteristics and time characteristics, the characteristic flattening layer performs flattening processing on the space-time characteristic tensor information according to the space-time characteristic tensor information to obtain characteristic vectors for life regression, the full-connection regression network performs characteristic transformation processing according to the characteristic vector information to obtain high-order characteristic representation related to the life degradation state of a spacecraft, and the life prediction layer performs regression prediction processing according to the high-order characteristic representation information to obtain a repeatable service life prediction value of the spacecraft. The time-space block comprises a learnable dynamic graph convolution sub-module and a time convolution sub-module (TCN), wherein the learnable dynamic graph convolution sub-module executes graph convolution operation according to an initial adjacent matrix, a sample specific dynamic matrix and a global learnable matrix of the TCN from an input layer or a previous stage of time-space block to aggregate space association information to obtain space characteristics reflecting the space dependence of a sensor, and the TCN