A telemetry data prediction method, apparatus, medium, and product

Abstract

The application discloses a kind of telemetry data prediction method, device, medium and product, it is related to spacecraft telemetry data prediction field.The method includes: obtaining the telemetry data corresponding to different variables of target spacecraft in each time in the time sequence to be measured, to obtain the real-time telemetry data sequence corresponding to each variable;Real-time telemetry data sequence corresponding to multiple variables is all input into trained multi-scale time convolution network, to obtain real-time time correlation feature;According to the real-time telemetry data sequence corresponding to multiple variables, based on similarity principle and trained multi-head attention mechanism, determine real-time graph structure feature;Real-time time correlation feature and real-time graph structure feature are all input into trained graph attention network, to obtain the telemetry prediction data corresponding to different variables of next time of current time.The application can simultaneously capture the time dependence of telemetry data and the spatial dependence between each variable, to further improve the prediction accuracy of spacecraft telemetry data.

Description

A method, apparatus, medium, and product for predicting telemetry data. Technical Field

[0001] This invention relates to the field of spacecraft telemetry data prediction, and in particular to a telemetry data prediction method, device, medium, and product. Background Technology

[0002] Telemetry data is a data stream received by the ground control center from the spacecraft in orbit. It includes measurements and status information from various sensors, such as attitude, temperature, and power, reflecting changes in the spacecraft's operating parameters and performance. Spacecraft are of immense value, and anomalies can lead to serious consequences and costly losses. Therefore, to reduce operational risks and ensure long-term stable operation, it is necessary to predict future telemetry data trends to detect potential anomalies and malfunctions in advance, thereby improving the spacecraft's operational safety and reliability.

[0003] Spacecraft typically consist of multiple subsystems, and their data exhibits characteristics such as high dimensionality, nonlinearity, and complexity. These data may contain correlations and dependencies, requiring powerful models to capture this information. Furthermore, telemetry data is time-series data, meaning it contains temporal correlations; past data may influence future predictions, necessitating time-series modeling. Current traditional telemetry data prediction methods, such as Autoregressive Integrated Moving Average (ARIMA), Support Vector Machines (SVMs), and Particle Swarm Optimization (PSO), often rely on domain knowledge to model data features, which is both complex and difficult to achieve satisfactory results. Moreover, these methods can only handle univariate data and cannot meet the requirements of multivariate telemetry data prediction. With the continuous development of deep learning, some deep learning-based prediction methods have been applied to telemetry data, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks. Although RNN and LSTM methods can handle multi-parameter prediction problems, they usually only consider time dependencies and rarely extract spatial dependencies between multiple variables in telemetry data. This makes it impossible for the model to discover and utilize the implicit relationships between parameters, resulting in low prediction accuracy of telemetry data. Summary of the Invention

[0004] The purpose of this invention is to provide a telemetry data prediction method, device, medium, and product that can simultaneously capture the temporal dependence of telemetry data and the spatial dependence between various variables, thereby improving the prediction accuracy of spacecraft telemetry data.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A telemetry data prediction method, comprising:

[0007] Determine the sequence of undetermined times within a sliding window before the current time using a preset step size;

[0008] Delete the first time in the undetermined time sequence and add the current time as the last element to the undetermined time sequence to obtain the time sequence to be tested;

[0009] Acquire telemetry data for different variables at each time point in the time sequence to be measured for the target spacecraft, and obtain the real-time telemetry data sequence for each variable;

[0010] The real-time telemetry data sequences corresponding to multiple variables are all input into the trained multi-scale temporal convolutional network to obtain real-time temporal correlation features;

[0011] Based on the real-time telemetry data sequences corresponding to multiple variables, the real-time graph structure features are determined according to the similarity principle and the trained multi-head attention mechanism.

[0012] The real-time temporal correlation features and the real-time graph structure features are both input into the trained graph attention network to obtain remote sensing prediction data corresponding to different variables in the next time step. The trained multi-scale temporal convolutional network, the trained multi-head attention mechanism, and the trained graph attention network are obtained by jointly training the multi-scale temporal convolutional network, the multi-head attention mechanism, and the graph attention network using historical telemetry data of the target spacecraft.

[0013] Optionally, the multi-scale temporal convolutional network includes:

[0014] A series of interconnected multi-scale convolutional layers, splicing layers, and multi-layer perceptron layers;

[0015] The multi-scale convolutional layer includes multiple parallel single-scale convolutional layers;

[0016] The kernel size of convolutional layers with different single-scale values ​​is different.

[0017] Optionally, based on the real-time telemetry data sequences corresponding to multiple variables, and using the principle of similarity and the trained multi-head attention mechanism, real-time graph structure features are determined, including:

[0018] Treat variables as nodes, and construct an embedding vector for each node based on the real-time telemetry data sequence corresponding to the node.

[0019] Based on multiple embedding vectors, the undirected influence score between any two nodes is determined using the similarity principle;

[0020] Multiple embedding vectors are input into the trained multi-head attention mechanism to determine the directed influence score between any two nodes;

[0021] Determine any node as the current node;

[0022] Determine the current node's own mixed influence score;

[0023] Identify all nodes other than the current node as potential neighbors of the current node;

[0024] Identify any undetermined neighbor node as the current undetermined neighbor node;

[0025] The sum of the undirected influence score of the current undetermined neighbor node on the current node and the directed influence score of the current undetermined neighbor node on the current node is the mixed influence score of the current undetermined neighbor node on the current node.

[0026] The difference between the mixed influence score of the current undetermined neighbor node on the current node and the mixed influence score of the current node itself is determined as the intermediate quantity corresponding to the current undetermined neighbor node.

[0027] Update the current undetermined neighbor node and return to the step "determine the undirected influence score of the current undetermined neighbor node on the current node, and the sum of the directed influence score of the current undetermined neighbor node on the current node, which is the mixed influence score of the current undetermined neighbor node on the current node" until all undetermined neighbor nodes of the current node are traversed, and the mixed influence score of all undetermined neighbor nodes on the current node and the intermediate quantity corresponding to all undetermined neighbor nodes are obtained.

[0028] The mean of the intermediate values ​​corresponding to all undetermined neighbor nodes is determined as the mixed influence score threshold.

[0029] Identify all undetermined neighbor nodes whose mixed influence score is greater than the mixed influence score threshold as the current node's neighbor nodes;

[0030] Update the current node and return to the step "determine the current node's own mixed influence score" until all nodes are traversed and the neighbor nodes of all nodes are obtained;

[0031] By connecting each node with its multiple neighboring nodes, we obtain the real-time graph structure features.

[0032] Optionally, determine the current node's own mixed influence score, including:

[0033] Determine the undirected influence score of the current node on the current node as the undirected influence score of the current node itself;

[0034] Determine the directed influence score of the current node on the current node as the directed influence score of the current node itself;

[0035] The sum of the current undirected influence score and the current directed influence score of a node is determined as the mixed influence score of the current node.

[0036] Optionally, before determining the sequence of time points to be determined within a sliding window with a preset step size before the current time, the method further includes:

[0037] Acquire historical telemetry data of the target spacecraft;

[0038] Define any historical moment as the current historical moment;

[0039] The sequence of undetermined historical moments is determined within a sliding window before the current historical moment using a preset step size.

[0040] Delete the first moment in the historical undetermined moment sequence and add the current historical moment as the last element to the historical undetermined moment sequence to obtain the current historical moment sequence;

[0041] Acquire telemetry data for different variables at each historical moment in the current historical time sequence of the target spacecraft, and obtain the historical telemetry data sequence for each variable;

[0042] Obtain telemetry data corresponding to different variables at the next time step from the current historical moment;

[0043] Using the historical telemetry data sequences corresponding to multiple variables at the current historical moment as historical inputs, and the telemetry data corresponding to different variables at the next moment as historical outputs, we construct historical data pairs corresponding to the current historical moment.

[0044] Update the current historical moment and return to the step "determine the sequence of undetermined historical moments within the sliding window before the current historical moment with a preset step size" until the preset number of iterations is reached, and obtain historical data pairs corresponding to multiple historical moments;

[0045] The historical data pairs corresponding to multiple historical moments are divided into test set and test set according to a preset ratio.

[0046] Optionally, after dividing the historical data pairs corresponding to multiple historical moments into training and test sets according to a preset ratio, the method further includes:

[0047] Multiple historical inputs from the test set are fed into a multi-scale temporal convolutional network to obtain multiple historical temporal correlation features.

[0048] Based on multiple historical inputs in the test set, multiple historical graph structural features are determined using the similarity principle and multi-head attention mechanism.

[0049] Using the historical time correlation features and historical graph structure features corresponding to the same historical data pair as inputs, and the historical outputs in the corresponding historical data pair as outputs, a multi-scale temporal convolutional network, a multi-head attention mechanism, and a graph attention network are jointly trained to obtain an undetermined multi-scale temporal convolutional network, an undetermined multi-head attention mechanism, and an undetermined graph attention network.

[0050] Optionally, after jointly training a multi-scale temporal convolutional network, a multi-head attention mechanism, and a graph attention network using the historical time correlation features and historical graph structure features corresponding to the same historical data pair as input, and the historical outputs in the corresponding historical data pair as output, to obtain the undetermined multi-scale temporal convolutional network, undetermined multi-head attention mechanism, and undetermined graph attention network, the following steps are also included:

[0051] The parameters of the undetermined multi-scale temporal convolutional network, the undetermined multi-head attention mechanism, and the undetermined graph attention network are tested and adjusted using the test set to obtain the trained multi-scale temporal convolutional network, the trained multi-head attention mechanism, and the trained graph attention network.

[0052] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the telemetry data prediction method described above.

[0053] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the telemetry data prediction method described above.

[0054] A computer program product includes a computer program that, when executed by a processor, implements the telemetry data prediction method described above.

[0055] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0056] This invention provides a telemetry data prediction method, device, medium, and product that acquires historical telemetry data of a target spacecraft; preprocesses the raw telemetry data to form training and testing sets that can be input into a prediction model; then inputs the training and testing sets into a multi-scale temporal convolutional network, a graph construction method, and a graph attention network to capture the temporal and spatial dependencies of the time-series data; finally, it uses the extracted spatiotemporal dependencies to predict the telemetry data. This method provides a more comprehensive and in-depth technical solution for spacecraft telemetry data prediction from both temporal and spatial dimensions. The multi-scale temporal convolutional network effectively captures correlations across different time ranges, while the graph construction method and graph attention network model the complex relationships between multiple variables in space, making the model more adaptable and accurate in prediction. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 is a flowchart of a telemetry data prediction method provided in Embodiment 1 of the present invention;

[0059] Figure 2 is a schematic diagram of a telemetry data prediction method provided in Embodiment 1 of the present invention;

[0060] Figure 3 is a schematic diagram of the multi-scale temporal convolutional network provided in Embodiment 1 of the present invention;

[0061] Figure 4 is a schematic diagram of the real-time graph structure feature determination method provided in Embodiment 1 of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The purpose of this invention is to provide a telemetry data prediction method, device, medium, and product that can simultaneously capture the temporal dependence of telemetry data and the spatial dependence between various variables, thereby improving the prediction accuracy of spacecraft telemetry data.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Example 1

[0066] As shown in Figure 1, a telemetry data prediction method in this embodiment includes:

[0067] Step 101: Determine the sequence of undetermined times within the sliding window before the current time using a preset step size.

[0068] Step 102: Delete the first time in the undetermined time sequence and add the current time as the last element to the undetermined time sequence to obtain the time sequence to be tested.

[0069] Step 103: Obtain telemetry data corresponding to different variables at each moment in the time sequence to be measured for the target spacecraft, and obtain the real-time telemetry data sequence corresponding to each variable.

[0070] Step 104: Input the real-time telemetry data sequences corresponding to multiple variables into the trained multi-scale temporal convolutional network to obtain real-time temporal correlation features.

[0071] Step 105: Based on the real-time telemetry data sequences corresponding to multiple variables, determine the real-time graph structure features based on the similarity principle and the trained multi-head attention mechanism.

[0072] Step 106: Input both the real-time temporal correlation features and the real-time graph structure features into the trained graph attention network to obtain the remote sensing prediction data corresponding to different variables at the next time step. The trained multi-scale temporal convolutional network, the trained multi-head attention mechanism, and the trained graph attention network are obtained by jointly training the multi-scale temporal convolutional network, the multi-head attention mechanism, and the graph attention network using historical telemetry data of the target spacecraft.

[0073] The multi-scale temporal convolutional network consists of sequentially connected multi-scale convolutional layers, splicing layers, and multi-layer perceptron layers. Each multi-scale convolutional layer comprises multiple parallel single-scale convolutional layers. The kernel sizes of the different single-scale convolutional layers vary.

[0074] Step 105 includes:

[0075] Step 105-1: Treat the variables as nodes and construct the embedding vector for each node based on the real-time telemetry data sequence corresponding to the node.

[0076] Step 105-2: Based on multiple embedding vectors, determine the undirected influence score between any two nodes using the similarity principle.

[0077] Step 105-3: Input multiple embedding vectors into the trained multi-head attention mechanism to determine the directed influence score between any two nodes.

[0078] Step 105-4: Determine any node as the current node.

[0079] Step 105-5: Determine the self-mixed influence score of the current node.

[0080] Step 105-6: Determine all nodes other than the current node as potential neighbor nodes of the current node.

[0081] Step 105-7: Determine any pending neighbor node as the current pending neighbor node.

[0082] Step 105-8: Determine the undirected influence score of the current undetermined neighbor node on the current node, and the sum of the directed influence score of the current undetermined neighbor node on the current node, which is the mixed influence score of the current undetermined neighbor node on the current node.

[0083] Step 105-9: Determine the difference between the mixed influence score of the current undetermined neighbor node on the current node and its own mixed influence score, which is the intermediate quantity corresponding to the current undetermined neighbor node.

[0084] Step 105-10: Update the current undetermined neighbor node and return to step 105-8 until all undetermined neighbor nodes of the current node are traversed, to obtain the mixed influence score of all undetermined neighbor nodes on the current node, and the intermediate quantity corresponding to all undetermined neighbor nodes.

[0085] Step 105-11: Determine the mean of the intermediate values ​​corresponding to all undetermined neighbor nodes as the mixed influence score threshold.

[0086] Step 105-12: Determine all undetermined neighbor nodes whose mixed influence score is greater than the mixed influence score threshold as the neighbor nodes of the current node.

[0087] Step 105-13: Update the current node and return to step 105-5 until all nodes have been traversed, obtaining the neighbor nodes of all nodes.

[0088] Steps 105-14: Connect each node with its corresponding multiple neighboring nodes to obtain the real-time graph structure features.

[0089] Step 105-5 includes:

[0090] Step 105-5-1: Determine the undirected influence score of the current node on the current node as the undirected influence score of the current node itself.

[0091] Step 105-5-2: Determine the directed influence score of the current node on the current node as the directed influence score of the current node itself.

[0092] Step 105-5-3: Determine the sum of the current undirected influence score and the current directed influence score of the node as the mixed influence score of the node.

[0093] Before step 101, the following is also included:

[0094] Step 107: Obtain historical telemetry data of the target spacecraft.

[0095] Step 108: Determine any historical moment as the current historical moment.

[0096] Step 109: Determine the sequence of undetermined historical moments within the sliding window before the current historical moment using a preset step size.

[0097] Step 1010: Delete the first moment in the historical undetermined moment sequence and add the current historical moment as the last element to the historical undetermined moment sequence to obtain the current historical moment sequence.

[0098] Step 1011: Obtain telemetry data corresponding to different variables at each historical moment in the current historical time sequence of the target spacecraft, and obtain the historical telemetry data sequence corresponding to each variable.

[0099] Step 1012: Obtain telemetry data corresponding to different variables at the next time step from the current historical moment.

[0100] Step 1013: Using the historical telemetry data sequences corresponding to multiple variables corresponding to the current historical time sequence as historical inputs, and the telemetry data corresponding to different variables at the next historical time as historical outputs, construct the historical data pairs corresponding to the current historical time.

[0101] Step 1014: Update the current historical moment and return to step 109 until the preset number of iterations is reached, and obtain historical data pairs corresponding to multiple historical moments.

[0102] Step 1015: Divide the historical data pairs corresponding to multiple historical moments into test set and test set according to the preset ratio.

[0103] Step 1016: Input multiple historical inputs from the test set into a multi-scale temporal convolutional network to obtain multiple historical temporal correlation features.

[0104] Step 1017: Based on the multiple historical inputs in the test set, determine multiple historical graph structural features using the similarity principle and multi-head attention mechanism.

[0105] Step 1018: Using the historical time correlation features and historical graph structure features corresponding to the same historical data pair as inputs, and the historical outputs in the corresponding historical data pair as outputs, jointly train the multi-scale temporal convolutional network, the multi-head attention mechanism, and the graph attention network to obtain the undetermined multi-scale temporal convolutional network, the undetermined multi-head attention mechanism, and the undetermined graph attention network.

[0106] Step 1019: Test and adjust the parameters of the undetermined multi-scale temporal convolutional network, the undetermined multi-head attention mechanism, and the undetermined graph attention network using the test set to obtain the trained multi-scale temporal convolutional network, the trained multi-head attention mechanism, and the trained graph attention network.

[0107] The telemetry data prediction method provided in this embodiment will be described in detail below with reference to Figure 2. This embodiment includes:

[0108] S1. Acquire raw telemetry data of the target spacecraft.

[0109] In step S1, the target spacecraft is the object for which telemetry data prediction is required, and may include, but is not limited to, artificial Earth satellites, probes, or space shuttles. The raw recorded telemetry data refers to unprocessed historical telemetry data.

[0110] S2. The original telemetry data is processed to remove outliers and missing values, resulting in processed telemetry data, which is then organized into N rows of time-series data.

[0111] In step S2, the original telemetry data may be affected by sensor damage or errors during data transmission, resulting in outliers and missing values. Therefore, preprocessing of the telemetry data in these original records is necessary. For outliers, this invention directly removes outlier data points from the original telemetry data. For missing values, this invention uses the K-nearest neighbor algorithm to fill in the missing data bits. Specifically, for each missing data point, the distance between the k similar data points before and after the historical data is calculated, and then the average of these k similar data points is used to estimate the value of the missing data point. After obtaining the telemetry data after the preprocessing operation, it is organized into N rows of time-series data, where N is a preset integer greater than or equal to 2, representing the total number of variables in the telemetry data. The N rows of time-series data can be, but are not limited to, represented as (X1, X2, ..., X...). t ,…,X T ), where X t Let N represent the N telemetry data points corresponding to time t, and T represent the total length of the N rows of time series data.

[0112] S3. Obtain multiple batches of sample data from the N rows of time-series data as follows: Starting from the first time point of the N rows of time-series data, extract the time-series data as model input data with a step size of 1 and a sliding window size of L. Simultaneously, use the N rows of time-series data at a single time point after each window as model output data, thus obtaining multiple batches of sample data containing the model input data and output data.

[0113] In step 3, L represents the length of the sliding window, which is the N rows of time-series data containing L time points; the step size represents the number of time points the sliding window moves in the direction of time growth (generally to the right) each time; the N rows of time-series data at a single time point after each sliding window are the label data of the corresponding sliding window. The N rows of time-series data contained in a single sliding window and the corresponding data labels constitute a single batch of sample data, and multiple such single batches of sample data constitute the multi-batch sample data.

[0114] S4. Divide the multiple batches of sample data into a training set and a test set in a 7:3 ratio.

[0115] In step 4, the training set is used to train the deep learning model. This involves continuously updating the model's parameters using data from the training set to better fit and understand the spatiotemporal characteristics of the telemetry data, and then saving the model's parameters. The test set primarily aims to evaluate the performance of the trained model on unseen data, observe the model's output, and ensure the model's reliability. The training and test sets are divided in a 7:3 ratio, with the training set's telemetry data having earlier temporal information than the test set's.

[0116] S5. The multi-scale temporal convolutional network is trained and tested sequentially using the training set and the test set to obtain the temporal correlation of telemetry time series data.

[0117] In step 5, the multi-scale temporal convolutional network is used to learn the temporal correlation of the telemetry data in the training and test sets, thereby capturing its temporal dependence.

[0118] The multi-scale temporal convolutional network proposed in this invention is a deep learning network specifically designed for processing temporal data. Its unique feature lies in its ability to apply convolutional operations at different time scales. By introducing multi-scale convolutional operations, the model can simultaneously focus on short-term, medium-term, and long-term changes in the data, thereby better capturing temporal patterns and dependencies. Specifically, the multi-scale temporal convolutional network uses three temporal convolutional networks with different kernel sizes. Smaller kernels capture short-term local patterns, medium-sized kernels capture medium-term patterns, and larger kernels help capture global patterns over a longer time span. This multi-scale convolutional operation makes the network more flexible and powerful in temporal data modeling. A schematic diagram is shown in Figure 3, and the equations are as follows:

[0119]

[0120] In the formula: X represents the input data; Z represents the output data after multiple temporal convolution operations; f 1×3 f 1×5 and f 1×7These represent three convolutional kernels of sizes 1×3, 1×5, and 1×7, respectively; * represents the convolution operation; concat is the concatenation operation, used to concatenate and truncate the input data after the three convolutional kernels; O represents the final output data of the multi-scale temporal convolutional network; MLP is the multilayer perceptron mapping operation, which, after the three convolutional operations, feeds the concatenated data Z into the multilayer perceptron to map and obtain the final multi-scale temporal convolutional output.

[0121] S6. The graph construction method is trained and tested sequentially using the training set and the test set to obtain the spatial correlation between telemetry time series data variables.

[0122] In step S6, the purpose of the graph construction method is to provide a suitable graph structure for the subsequent graph attention network, ensuring that the network can fully utilize the neighbor node information associated with its own nodes. The graph structure refers to the adjacency matrix of the graph, which contains the neighbor node information of each node. Unlike previous methods that only considered a single graph type, this invention considers both undirected and directed graphs. This design not only considers the correlation between nodes but also the information flow between nodes, thus compensating for the shortcomings and deficiencies of a single graph type. Specifically, this invention uses cosine similarity to capture the correlation between variables and uses a multi-head attention mechanism to capture the information flow between variables. This comprehensive approach helps to more comprehensively model the complex relationships between multiple variables, improving the performance and generalization ability of the graph attention network. A schematic diagram of the graph construction method is shown in Figure 4, and the detailed process is as follows:

[0123] Assign a randomly initialized embedding vector e to each variable or node. i Where i∈{1,2,…,N}. The embedding vectors are then used with cosine similarity and multi-head attention mechanisms to obtain undirected influence scores and directed influence scores, respectively, as shown in the following formulas:

[0124]

[0125] In the formula: γ is the undirected influence score, α is the directed influence score, and α is the mixed influence score obtained by adding the undirected influence score and the directed influence score. e is the transpose of the embedding vector of node i. j Let be the embedding vector of node j; in the second fraction, the numerator represents the inner product of the two vectors, and the denominator represents the product of their vector magnitudes; Q, K, V, and W are the trainable parameter matrices; e is the matrix form composed of the embedding vectors of all nodes; Attention is the operation of the multi-head attention mechanism. The larger the mixed influence score, the greater the potential relationship and information flow between the two nodes.

[0126] From the mixed influence scores, an element of the adjacency matrix of the final graph structure can be defined as follows:

[0127]

[0128] In the formula: α ii For the mixed influence score of its own node, α ji μ is the mixed influence score of node j on node i. i α represents the mean of the mixed influence scores of all other nodes on node i after removing its own mixed influence score; k As a variable, α can be substituted. ii or α ji A ji =1 indicates that there is an edge connecting node i and node j.

[0129] First, the mixed influence score of each node is removed. Then, the average mixed influence score of the remaining nodes on each node is calculated. Finally, nodes whose remaining influence scores are greater than or equal to the average are selected as their neighbors. After each node is processed in this way, an adjacency matrix is ​​constructed, which is the final learned graph structure, representing the network of relationships between nodes.

[0130] S7. The training set, the test set, and the telemetry data obtained in S5 and S6 are used as spatiotemporal inputs to the graph attention network for training and testing in sequence, and the parameters of the prediction model that passes the test are obtained and saved.

[0131] In step 7, the graph attention network first uses the graph structure obtained in step S6, then uses the output of the multi-scale temporal convolutional network described in step S5 as the input to the graph attention network to update the node embedding representation. Finally, the updated node embedding representation is fed into a multilayer perceptron to obtain the final predicted telemetry data output. The specific prediction implementation of the graph attention network is as follows:

[0132] Calculate the attention coefficient between node i and its neighbor node j one by one:

[0133] β ij =a([Wh i ||Wh j ]), j∈M i .

[0134] In the formula: W is the parameter matrix to be learned, h i h is the embedding representation of node i. j Let M be the embedding representation of node i's neighbor node j. i Let Wh be the set of neighboring nodes of node i. iThis indicates that the shared parameter W is used to represent h. i A linear transformation is performed; [·||·] represents the concatenation of the transformed results from nodes i and j; a is a parameter matrix used to map the concatenated high-dimensional features onto a real number, yielding the attention coefficients. The attention coefficients are then normalized.

[0135]

[0136] Where: β ij The attention coefficients between nodes i and j are represented by: LeakyReLU, an activation function that performs a non-linear transformation on the attention coefficients; and softmax, a normalization mapping function that normalizes the attention coefficients. The normalized attention coefficients are then weighted and summed to update the node embedding representation.

[0137]

[0138] Where: π ij The normalized attention coefficients between nodes i and j are represented by σ; σ is the activation function; h′ i Let be the embedding representation of node i after one iteration of the graph attention network. Finally, the updated embedding representation is input into a multilayer perceptron to obtain the final telemetry data prediction output.

[0139] Y=MLP(h′1,h′2,…,h′ N ).

[0140] In the formula, MLP represents multilayer perceptron operation; Y is the final predicted telemetry data obtained after model inference. After obtaining the predicted data, this invention uses the mean absolute error as the loss function of the model.

[0141] S8. Input the latest sliding window data of the target spacecraft into the prediction model and load the saved model parameters, and output the telemetry data of the spacecraft at the next moment of the latest sliding window.

[0142] In step 8, the latest sliding window data of the spacecraft is preprocessed according to step S2, and the preprocessed data is used as the input data for the model. The model parameters are the prediction model parameters saved after the test in step S7.

[0143] Example 2

[0144] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a telemetry data prediction method as described in Embodiment 1.

[0145] Example 3

[0146] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a telemetry data prediction method as described in Embodiment 1.

[0147] Example 4

[0148] A computer program product includes a computer program that, when executed by a processor, implements a telemetry data prediction method as described in Embodiment 1.

[0149] Example 5

[0150] A computer device, which may be a database, has an internal structure as shown in Figure 3. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores pending transactions. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a telemetry data prediction method as described in Embodiment 1.

[0151] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0152] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0153] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting telemetry data, characterized in that, include: Determine the sequence of undetermined times within a sliding window before the current time using a preset step size; The first moment in the undetermined time sequence is deleted, and the current moment is added as the last element to the undetermined time sequence to obtain the undetermined time sequence; the telemetry data corresponding to different variables of the target spacecraft at each moment in the undetermined time sequence are obtained to obtain the real-time telemetry data sequence corresponding to each variable; the real-time telemetry data sequences corresponding to multiple variables are input into the trained multi-scale temporal convolutional network to obtain the real-time temporal correlation features. Based on real-time telemetry data sequences corresponding to multiple variables, and using the similarity principle and a trained multi-head attention mechanism, the real-time graph structure features are determined, including: treating variables as nodes and constructing an embedding vector for each node based on the corresponding real-time telemetry data sequence; determining the undirected influence score between any two nodes using the similarity principle based on multiple embedding vectors; inputting multiple embedding vectors into the trained multi-head attention mechanism to determine the directed influence score between any two nodes; determining any node as the current node; determining the current node's own mixed influence score; and determining all nodes except the current node. The process involves identifying potential neighbor nodes for the current node; determining any potential neighbor node as the current potential neighbor node; calculating the undirected influence score of the current potential neighbor node on the current node, and the sum of its directed influence score and undirected influence score, as the mixed influence score of the current potential neighbor node on the current node; determining the difference between the mixed influence score of the current potential neighbor node on the current node and its own mixed influence score, as the intermediate value corresponding to the current potential neighbor node; updating the current potential neighbor node and returning to step "determine the undirected influence score of the current potential neighbor node on the current node, and the sum of its directed influence score and undirected influence score of the current potential neighbor node on the current node". The sum of the directed influence scores of each node on the current node is the mixed influence score of the current undetermined neighbor nodes on the current node. This process continues until all undetermined neighbor nodes of the current node are traversed, obtaining the mixed influence scores of all undetermined neighbor nodes on the current node, as well as the intermediate values ​​corresponding to all undetermined neighbor nodes. The mean of the intermediate values ​​corresponding to all undetermined neighbor nodes is determined as the mixed influence score threshold. All undetermined neighbor nodes whose mixed influence scores on the current node are greater than the mixed influence score threshold are determined as the neighbors of the current node. The current node is updated and the process returns to the step "determine the mixed influence score of the current node itself" until all nodes are traversed, obtaining the neighbors of all nodes. Each node is connected to its corresponding multiple neighbor nodes to obtain the real-time graph structure features. The real-time temporal correlation features and the real-time graph structure features are both input into the trained graph attention network to obtain the remote sensing prediction data corresponding to different variables at the next time step. The trained multi-scale temporal convolutional network, the trained multi-head attention mechanism, and the trained graph attention network are obtained by jointly training the multi-scale temporal convolutional network, the multi-head attention mechanism, and the graph attention network using historical telemetry data of the target spacecraft.

2. The telemetry data prediction method according to claim 1, characterized in that, The multi-scale temporal convolutional network includes: a multi-scale convolutional layer, a splicing layer, and a multi-layer perceptron layer connected in sequence; the multi-scale convolutional layer includes multiple parallel single-scale convolutional layers; the convolutional kernels of different single-scale convolutional layers have different sizes.

3. The telemetry data prediction method according to claim 1, characterized in that, Determine the current node's self-mixed influence score, including: determining the current node's undirected influence score on the current node as its own undirected influence score; determining the current node's directed influence score on the current node as its own directed influence score; and determining the sum of the current node's undirected influence score and its own directed influence score as the current node's self-mixed influence score.

4. The telemetry data prediction method according to claim 1, characterized in that, Before determining the sequence of undetermined time points within a sliding window with a preset step size, the process includes: acquiring historical telemetry data of the target spacecraft; determining any historical time point as the current historical time point; determining the sequence of undetermined historical time points within a sliding window with a preset step size; deleting the first time point in the sequence of undetermined historical time points and adding the current historical time point as the last element to the sequence of undetermined historical time points, thus obtaining the current historical time point sequence; acquiring telemetry data corresponding to different variables of the target spacecraft at each historical time point in the current historical time point sequence, thus obtaining the historical telemetry data sequence corresponding to each variable; acquiring telemetry data corresponding to different variables at the next time point of the current historical time; using the historical telemetry data sequences corresponding to multiple variables of the current historical time point sequence as historical inputs and the telemetry data corresponding to different variables at the next time point of the current historical time as historical outputs, constructing historical data pairs corresponding to the current historical time point; updating the current historical time point and returning to the step "determining the sequence of undetermined historical time points within a sliding window with a preset step size", until the iteration reaches a preset number, resulting in multiple historical data pairs corresponding to multiple historical time points; and dividing the multiple historical data pairs corresponding to multiple historical time points into a test set and a test group according to a preset ratio.

5. The telemetry data prediction method according to claim 4, characterized in that, After dividing the historical data pairs corresponding to multiple historical moments into training and testing sets according to a preset ratio, the process further includes: inputting multiple historical inputs from the testing set into a multi-scale temporal convolutional network to obtain multiple historical temporal correlation features; determining multiple historical graph structure features based on the similarity principle and multi-head attention mechanism according to the multiple historical inputs from the testing set; and jointly training the multi-scale temporal convolutional network, multi-head attention mechanism, and graph attention network with the historical temporal correlation features and historical graph structure features corresponding to the same historical data pair as inputs and the historical outputs in the corresponding historical data pair as outputs to obtain the undetermined multi-scale temporal convolutional network, undetermined multi-head attention mechanism, and undetermined graph attention network.

6. The telemetry data prediction method according to claim 5, characterized in that, After jointly training a multi-scale temporal convolutional network, a multi-head attention mechanism, and a graph attention network using the historical time correlation features and historical graph structure features corresponding to the same historical data pair as inputs and the historical outputs in the corresponding historical data pair as outputs, to obtain the undetermined multi-scale temporal convolutional network, undetermined multi-head attention mechanism, and undetermined graph attention network, the training further includes: testing and adjusting the parameters of the undetermined multi-scale temporal convolutional network, undetermined multi-head attention mechanism, and undetermined graph attention network using the test set, to obtain the trained multi-scale temporal convolutional network, trained multi-head attention mechanism, and trained graph attention network.

7. A computer device, comprising: The memory and processor contain a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement a telemetry data prediction method according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a telemetry data prediction method according to any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements a telemetry data prediction method according to any one of claims 1-6.

Images