Spacecraft reusable service life prediction method based on learnable dynamic graph neural network
By using a learnable dynamic graph neural network-based method to dynamically adjust the correlation between spacecraft sensors, the problem of predefined graph structures being unable to adapt and adjust in existing technologies is solved, enabling more accurate spacecraft lifetime prediction and reusability determination.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-01-24
- Publication Date
- 2026-05-08
AI Technical Summary
Existing spacecraft lifetime prediction technologies cannot adaptively adjust the correlation between sensors according to changes in different operating conditions, different reusability stages, and degradation levels of spacecraft, resulting in limited prediction accuracy and generalization ability.
By employing a learnable dynamic graph neural network-based approach, a training set is generated to process time-series data from multiple sensors, constructing an initial adjacency matrix, a globally learnable matrix, and a sample-specific dynamic matrix. The graph structure is then dynamically adjusted to achieve end-to-end lifetime prediction.
It achieves more accurate and stronger lifetime prediction capabilities, reduces space modeling errors caused by fixed topology assumptions, and improves the reliability and engineering applicability of spacecraft reusability determination.
Smart Images
Figure CN121997455A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the aerospace field, specifically a method for predicting the repeatable lifespan of spacecraft based on a learnable dynamic graph neural network. Background Technology
[0002] In spacecraft health management, sensors are typically deployed on multiple key subsystems to collect operational data such as temperature, pressure, vibration, and electrical parameters in real time. This multi-source sensor data records the spacecraft's operational status and degradation process in time-series format, while also implicitly revealing the coupling relationships between different sensors. Existing lifetime prediction techniques based on graph neural networks usually employ predefined static graph structures, failing to adaptively adjust the correlations between sensors according to changes in different spacecraft operating conditions, reuse stages, and degradation levels. Furthermore, lifetime assessment techniques based on the coupling of image fields and physical fields lack the means to dynamically adjust the correlation structure between features based on changes in equipment operating conditions or degradation stages, thus failing to characterize the dynamic coupling relationships that arise as the multi-sensor system evolves with changing states. Summary of the Invention
[0003] This invention addresses the shortcomings of existing lifetime prediction technologies, which rely on predefined adjacency matrices, cannot adaptively adjust based on input samples or degradation stages, and struggle to capture dynamic relationships between sensors, resulting in limited accuracy and generalization ability in lifetime prediction. It proposes a spacecraft repeatable lifetime prediction method based on a learnable dynamic graph neural network. This method can learn and dynamically adjust the graph structure end-to-end, solving the problem of inflexibility associated with fixed graph structures and achieving more accurate and generalized lifetime prediction.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a method for predicting the reusable lifespan of spacecraft based on a learnable dynamic graph neural network. In the offline phase, time-series data from multiple sensors are collected and a training set is generated to train a spatiotemporal feature prediction network containing several spatiotemporal blocks. In the online phase, the trained spatiotemporal feature prediction network is used for real-time lifespan prediction.
[0006] The training set is obtained by cleaning and preprocessing multi-sensor time-series data.
[0007] The spatiotemporal feature prediction network comprises: an input layer, several cascaded spatiotemporal blocks, a feature flattening layer, a fully connected regression network, and a lifetime prediction layer. Specifically: the input layer standardizes the time-series data from multiple sensors and segments it using a sliding window method to obtain a multi-sensor time-series input matrix; each cascaded spatiotemporal block performs spatial relationship modeling and temporal dependency modeling based on the multi-sensor time-series input matrix from the input layer or the previous spatiotemporal block, resulting in a spatiotemporal feature tensor that integrates spatial and temporal features; the feature flattening layer flattens the spatiotemporal feature tensor in both the time and sensor dimensions to obtain feature vectors for lifetime regression; the fully connected regression network performs feature transformation through nonlinear mapping based on the feature vector information to obtain a high-order feature representation related to the spacecraft's lifetime degradation state; and the lifetime prediction layer performs regression prediction based on the high-order feature representation to obtain a predicted repeatable lifetime value for the spacecraft.
[0008] The spatiotemporal block includes a learnable dynamic graph convolution submodule and a temporal convolution submodule (TCN). The learnable dynamic graph convolution submodule performs graph convolution operations based on the initial adjacency matrix, sample-specific dynamic matrix, and global learnable matrix from the TCN of the input layer or the previous spatiotemporal block to aggregate spatial correlation information, obtaining spatial features reflecting sensor spatial dependencies. The TCN, based on these spatial features, models the degradation and evolution characteristics in the time series using causal dilated temporal convolution, obtaining temporal features containing temporal dependencies, specifically: , Where: k is the layer index of the spatiotemporal block, The multi-sensor spatiotemporal feature tensor output by the (k-1)th spatiotemporal block is used as the input feature of the current spatiotemporal block. The intermediate spatial feature tensor is obtained after processing by the learnable dynamic graph convolution submodule in the k-th spatiotemporal block. This is the spatiotemporal feature tensor output after further processing by the TCN in the k-th spatiotemporal block. For learnable dynamic graph convolution submodules, This is a temporal convolutional network submodule.
[0009] The initial adjacency matrix is constructed by calculating the initial connection strength between sensor nodes based on the statistical correlation information of historical operating data from multiple sensors.
[0010] The globally learnable matrix is obtained by adaptively optimizing the adjacency relationship based on the parameter update information obtained by backpropagation during training, thus obtaining a globally learnable matrix used to represent global spatial dependencies.
[0011] The sample-specific dynamic matrix is obtained by generating dynamic connection relationships between nodes through embedding mapping and correlation calculation, based on the initial adjacency matrix of the TCN from the input layer or the previous spatiotemporal block.
[0012] The training described uses mean squared error (MSE) as the loss function to optimize all model weights in an end-to-end manner.
[0013] In the online phase, the predicted reusable lifespan of the spacecraft is compared with a pre-set flight lifespan threshold, wherein the flight lifespan threshold is determined based on the operating conditions, safety margin, and maintenance strategy for the next flight mission.
[0014] Technical effect
[0015] This invention addresses the scenario of multi-sensor lifetime prediction for spacecraft by proposing a composite adjacency matrix construction and fusion mechanism, which consists of an initial adjacency matrix, a globally learnable matrix, and a sample-specific dynamic matrix. The composite adjacency matrix is then introduced into a spatiotemporal feature prediction network to adaptively adjust the topological relationships between sensors during model training and inference as the spacecraft's operating conditions and degradation stages change, thereby enabling the modeling of the dynamic spatial dependence characteristics of the multi-sensor system. Attached Figure Description
[0016] Figure 1 This is a flowchart of the present invention;
[0017] Figure 2 This is a schematic diagram of the system in the embodiment;
[0018] Figure 3 This is a schematic diagram illustrating the effect of an example. Detailed Implementation
[0019] like Figure 1 As shown in the figure, this embodiment relates to a method for predicting the reusable lifespan of a spacecraft based on a learnable dynamic graph neural network, including:
[0020] Step 1: Preprocessing of multi-sensor time-series data.
[0021] The aforementioned time-series data refers to time-series data collected synchronously by N sensors during the operation of the spacecraft, with each sensor generating a measurement value at each time point.
[0022] The preprocessing mentioned refers to: after performing Z-score normalization on the data from each sensor, obtaining the input matrix using a sliding window. Specifically: ,in: For window length, For the real number field, For the raw data of the i-th sensor at timestamp t, and Let be the mean and standard deviation of the i-th sensor data in the training set, respectively. A sliding window strategy is used to construct the model input.
[0023] Step 2 involves modeling spatial relationships using a learnable dynamic graph convolution approach, specifically including:
[0024] 2.1 Calculation of the initial static adjacency matrix based on sensors Matrix elements This represents the initial connection strength between nodes i and j.
[0025] The initial static adjacency matrix is calculated using the cosine similarity method or the Pearson correlation coefficient.
[0026] 2.2 After constructing the adjacency matrix based on the initial static adjacency matrix, a dynamic graph structure is generated;
[0027] The adjacency matrix includes: a globally learnable matrix. This is used to optimize dynamic graph convolutional models during training via backpropagation, capturing hidden space dependencies that are prevalent in the dataset but not reflected in the initial graph, and generating sample-specific dynamic matrices in real time from the input matrix X after preprocessing in step 1 using a lightweight embedding function. .
[0028] The sample-specific dynamic matrix is obtained as follows: after mapping node features to the embedding space through a linear layer, the correlation between nodes is calculated to obtain the sample-specific dynamic matrix of the current input sample. ,in: , , The weight matrix is a learnable matrix. As a bias, the Softmax function is used to normalize the correlation strength.
[0029] 2.3 Sample-Specific Dynamic Matrix Performing graph convolution operations to aggregate spatial information enables the model to adaptively learn and optimize the topological relationships between sensors, specifically: ,in: For the spatial features of the l-th layer, This is the feature dimension of this layer. This is the trainable weight matrix for this layer.
[0030] Step 3, extract temporal dependencies through a temporal convolutional network: In order to capture the temporal dynamics of device degradation, the spatial features H obtained in Step 2 are input into a temporal convolutional network (TCN), and causal dilated convolution is used to capture long-term dependencies and extract key temporal degradation patterns from the sequence.
[0031] Step 4, through, as Figure 2 The spatiotemporal feature prediction network shown, comprising K cascaded spatiotemporal blocks, performs deep fusion of the features obtained in step 3 to obtain a high-order fused feature tensor. .
[0032] Step 5: Flatten the high-order fusion feature tensor obtained in Step 4 in both time and space dimensions, and then use a fully connected regression network to predict the repeatable lifetime value. Specifically: ,in: Indicates the flattening operation. and These are the weights and biases of a fully connected regression network.
[0033] The fully connected regression network consists of two fully connected layers and uses the ReLU activation function.
[0034] The spatiotemporal feature prediction network is trained using mean squared error (MSE) as the loss function and optimizes all model weights in an end-to-end manner, including graph convolution weights, TCN weights, fully connected layer weights, and learnable graph structure parameters B and embedding function parameters for generating C.
[0035] The training described uses the following loss function. Where: n is the number of samples, This is a predicted value for repeatable service life. This represents the actual value of the repeatable service life.
[0036] The training process preferably uses the Adam optimizer.
[0037] Preferably, the predicted reusable service life value obtained in step 5 is compared with a pre-set flight service life threshold: when the predicted value is greater than or equal to the threshold, the spacecraft is determined to meet the service life requirements for the next flight mission, and a reusable determination result is obtained; when the predicted value is less than the threshold, the spacecraft is determined to not meet the service life requirements for the next flight mission, and a non-reusable determination result is obtained.
[0038] The flight life threshold is determined based on the operating conditions, safety margin, and maintenance strategy for the next flight mission.
[0039] Through practical application experiments, this invention's spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network was run using the PyTorch framework on an NVIDIA RTX 4060Ti GPU hardware environment. The dataset used was the C-MAPSS aero-engine dataset (FD002 subset, containing 21 sensor data under six operating conditions; sensor channels S1, S5, S6, S10, S16, S18, and S19 with insignificant changes were first removed, retaining 14 key sensor signals). The experimental process included: reading training and test data from the FD002 subset, performing Z-score normalization and sliding window (T=60, step size=1) preprocessing; calculating the initial adjacency matrix A based on cosine similarity; generating C through an embedding layer and fusing the results. The model consists of three cascaded spatiotemporal blocks, with three stacked TCN layers (expansion rates 1, 2, and 4). Training uses MSE loss and the Adam optimizer (batch size 128, learning rate 0.02, step decay). Experimental data shows a root mean square error (RMSE) of 14.35 on the test set, with the true and predicted values compared to... Figure 3 As shown, the overlap is high.
[0040] Compared with existing technologies, this invention introduces a composite adjacency matrix into the lifetime prediction process, enabling the model to dynamically characterize the evolution of spatial relationships in a multi-sensor system and reducing spatial modeling errors caused by fixed topological assumptions. Combined with a spatiotemporal block cascade feature extraction structure, it achieves joint modeling of long-term degradation trends and spatial coupling characteristics. In spacecraft lifetime prediction missions, compared with lifetime prediction methods using static adjacency structures, it can obtain more stable and accurate lifetime prediction results, thereby improving the reliability and engineering applicability of spacecraft reusability determination, reducing reliance on human experience, and possessing significant engineering application value.
[0041] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for predicting the repeatable lifespan of a spacecraft based on a learnable dynamic graph neural network, characterized in that, In the offline phase, time-series data from multiple sensors are collected and a training set is generated to train a spatiotemporal feature prediction network containing several spatiotemporal blocks; in the online phase, the trained spatiotemporal feature prediction network is used for real-time lifetime prediction. The spatiotemporal feature prediction network includes: an input layer, K cascaded spatiotemporal blocks, a feature flattening layer, a fully connected regression network, and a lifetime prediction layer.
2. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 1, characterized in that, The input layer standardizes the data from each sensor and segments it using a sliding window method based on the time-series data collected synchronously from multiple sensors to obtain a multi-sensor time-series input matrix. Each cascaded spatiotemporal block performs spatial relationship modeling and temporal dependency modeling based on the multi-sensor time-series input matrix from the input layer or the previous spatiotemporal block, based on the dynamically changing sensor correlation relationship, to obtain a spatiotemporal feature tensor that integrates spatial and temporal features. The feature flattening layer flattens the spatiotemporal feature tensor information in the time and sensor dimensions to obtain feature vectors for lifetime regression. The fully connected regression network performs feature transformation processing through nonlinear mapping based on the feature vector information to obtain high-order feature representations related to the spacecraft's lifetime degradation state. The lifetime prediction layer performs regression prediction processing based on the high-order feature representation information to obtain the predicted value of the spacecraft's repeatable lifespan.
3. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 1, characterized in that, The spatiotemporal block includes a learnable dynamic graph convolution submodule and a temporal convolution submodule (TCN). The learnable dynamic graph convolution submodule performs graph convolution operations to aggregate spatial correlation information based on the initial adjacency matrix, sample-specific dynamic matrix, and global learnable matrix from the TCN of the input layer or the previous spatiotemporal block, thereby obtaining spatial features that reflect the spatial dependencies of the sensors. Based on the spatial feature information, the TCN uses causal dilated temporal convolution to model the degradation and evolution characteristics in the time series, thereby obtaining temporal features containing temporal dependencies.
4. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 3, characterized in that, The initial adjacency matrix is constructed by calculating the initial connection strength between sensor nodes based on the statistical correlation information of historical operating data from multiple sensors. The globally learnable matrix is obtained by adaptively optimizing the adjacency relationship based on the parameter update information obtained by backpropagation during training, thus obtaining a globally learnable matrix used to represent global spatial dependencies. The sample-specific dynamic matrix is obtained by generating dynamic connection relationships between nodes through embedding mapping and correlation calculation, based on the initial adjacency matrix of the TCN from the input layer or the previous spatiotemporal block.
5. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 3 or 4, characterized in that, The spatiotemporal feature tensor output after further processing by the TCN in the k-th spatiotemporal block, , Where: k is the layer index of the spatiotemporal block, The multi-sensor spatiotemporal feature tensor output by the (k-1)th spatiotemporal block is used as the input feature of the current spatiotemporal block. The intermediate spatial feature tensor is obtained after processing by the learnable dynamic graph convolution submodule in the k-th spatiotemporal block. For learnable dynamic graph convolution submodules, This is a temporal convolutional network submodule; The training described uses mean squared error (MSE) as the loss function to optimize all model weights in an end-to-end manner.
6. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 1, characterized in that, In the online phase, the predicted reusable lifespan of the spacecraft is compared with a pre-set flight lifespan threshold, wherein the flight lifespan threshold is determined based on the operating conditions, safety margin, and maintenance strategy for the next flight mission.
7. The spacecraft repeatable lifespan prediction method based on a learnable dynamic graph neural network according to claim 1, characterized in that, The fully connected regression network, based on the flattened high-order fusion feature tensor in the time and space dimensions, predicts the repeatable lifetime value, specifically as follows: ,in: Indicates the flattening operation. and For weights and biases.