A self-adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method

Abstract

The application provides a kind of adaptive multi-task balance parallel satellite constellation telemetry data monitoring method, comprising: step 1: obtaining satellite constellation telemetry time series data;Step 2: preprocessing satellite constellation telemetry time series data;Step 3: constructing satellite constellation telemetry data prediction model;Step 4: training satellite constellation telemetry data prediction model based on preprocessed satellite constellation telemetry time series data;Step 5: obtaining the predicted value of each telemetry channel based on the trained satellite constellation telemetry data prediction model, and real-time monitoring satellite constellation telemetry data.The application adopts the method of shared feature extraction combined with parallel prediction of each telemetry channel of different models of satellites in satellite constellation to fuse multi-dimensional parameter information, so that the models of each telemetry channel complement and enhance each other in the training process, and improve the overall data monitoring effect.

Description

Technical Field

[0001] This invention relates to an adaptive, multi-task balanced, parallel satellite constellation telemetry data monitoring method, belonging to the field of aerospace technology. Background Technology

[0002] Telemetry data monitoring is a hot research topic in the field of satellite status assessment and management. With the increasing maturity of microsatellite manufacturing and launch, their costs have been greatly reduced. Satellite constellations composed of a large number of satellites have become a focus of development due to their advantages such as wide coverage, low latency, and high reliability. Satellites in a constellation network need to work together to complete a specific task, which may involve multiple links such as signal detection, area positioning, and information acquisition. Therefore, the structure of a satellite constellation is highly coupled and locally heterogeneous. In addition, the satellites themselves have complex operating modes, high coupling between internal systems, and rich telemetry channel parameters. However, in actual scenarios, the computing resources for monitoring telemetry parameters are often limited. Therefore, methods for capturing shared information between telemetry channels and realizing the early reuse of monitoring tasks of each telemetry channel to simultaneously monitor telemetry channels of multi-dimensional satellite constellations have become a research hotspot.

[0003] In existing technologies, the China Academy of Space Technology (CN115575989) uses Long Short-Term Memory (LSTM) networks to perform time-series analysis on telemetry data from multiple channels, learning the temporal dependencies between telemetry data to predict future telemetry data patterns. This lays the groundwork for the detection of anomalies at the individual satellite or subsystem level. However, while LSTM networks can capture dynamic evolution information across multiple time dimensions, they are deficient in capturing the topological relationships of satellite constellation networks and the interrelationships of telemetry data between various satellite components. Furthermore, they struggle to account for the evolutionary characteristics of different satellites and components with varying attributes, making it difficult to achieve refined modeling of the operational status of satellite constellation components. The technical solution of this invention employs a multi-component-level decoder architecture to achieve fine-grained monitoring of satellite telemetry data. With a relatively small increase in parameters, it enables parallel dynamic capture of telemetry data change patterns in various sub-parts and types of components within the constellation, paving the way for subsequent implementation of... Precise anomaly localization at the component level provides a foundation for early detection. Nanjing University of Aeronautics and Astronautics (CN117688496) uses a causal-automatic learning graph neural network structure to simultaneously monitor multi-dimensional telemetry channels. However, it does not consider the learning difficulty of different satellite types and channels within the constellation, which may lead to the model favoring easier tasks, resulting in an imbalance in the monitoring performance of different telemetry channels. The technical solution of this invention uses a variance scoring ratio to characterize the learning degree of prediction tasks for each channel of each satellite type. It adaptively optimizes the model's learning degree for prediction tasks of each telemetry channel of each satellite type, slows down the learning speed of channels that are performing well at the current stage, and speeds up the learning speed of channels that are performing poorly at the current stage. This avoids performance degradation on some difficult telemetry channels due to excessive favoritism towards a certain easier telemetry channel, thereby improving the overall performance of the model across the entire satellite constellation. Summary of the Invention

[0004] To address the technical problem of unbalanced monitoring performance of multiple telemetry channels in heterogeneous satellite constellations in the prior art, this invention proposes an adaptive, multi-task balanced, parallel satellite constellation telemetry data monitoring method.

[0005] The technical solution adopted by this invention to solve the above problems is as follows: This invention proposes an adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method, comprising:

[0006] Step 1: Acquire satellite constellation telemetry time series data;

[0007] Step 2: Preprocess the satellite constellation telemetry time series data;

[0008] Step 3: Construct a satellite constellation telemetry data prediction model;

[0009] Step 4: Train the satellite constellation telemetry data prediction model based on the preprocessed satellite constellation telemetry time series data;

[0010] Step 5: Obtain the predicted value of each telemetry channel based on the trained satellite constellation telemetry data prediction model, and monitor the satellite constellation telemetry data in real time.

[0011] Optionally, step 2, which involves preprocessing the satellite constellation telemetry time series data, includes:

[0012] Step 2.1: Divide the satellite constellation telemetry time series data into a training telemetry dataset and a test telemetry dataset, with the ratio of the training set to the test set being 8:2;

[0013] Step 2.2: Standardize the training telemetry dataset by channel;

[0014] The expression for standardization is:

[0015]

[0016] In formula (1), For the training telemetry data of telemetry channel c at time t1, μ(X) trainc ) represents the mean of all training telemetry data for telemetry channel c, σ(X) trainc ) represents the standard deviation of all training telemetry data for telemetry channel c.

[0017] Optionally, the training telemetry dataset and test telemetry dataset after partitioning in step 2.1 include X train Y train , X test Y test and

[0018] The expressions for the training telemetry dataset and the test telemetry dataset are as follows:

[0019]

[0020]

[0021] In formulas (2)-(9), For the training telemetry data of telemetry channel c at time t1, For the test telemetry data of telemetry channel c at time t2, X train This refers to the training telemetry data after being segmented into windows of length w, where w ranges from 15 to 35. To train the telemetry data from time i (inclusive) to time w+i (exclusive), Y train For the corresponding X train In each window, the actual training telemetry data is predicted p time steps backward, where p ranges from 1 to 20. To train the telemetry data from time w+i (inclusive) to time w+p+i (exclusive), X test The data was segmented into windows of length w for testing telemetry data. To test the data values ​​of Y from time j (inclusive) to time w+j (exclusive) in the telemetry data, test For the corresponding X test In each window, the actual training telemetry data is predicted p time steps backward. To test the data values ​​from time w+j (inclusive) to time w+p+j (exclusive) in the telemetry data, T1 is the time span of the training telemetry data, ranging from 200 to 20000, and T2 is the time span of the test telemetry data, ranging from 200 to 20000.

[0022] Optionally, the satellite constellation telemetry data prediction model in step 3 includes a multi-channel shared layer and a mission-specific layer;

[0023] The multi-channel sharing layer is used to characterize the shared information between each telemetry channel and capture the correlation and commonalities between different tasks;

[0024] The mission-specific layer is used to perform fine-grained modeling of the characteristics of each telemetry channel in the satellite constellation.

[0025] Optionally, step 3, which involves constructing a satellite constellation telemetry data prediction model, includes:

[0026] Step 3.1: Use a single encoder to characterize the shared information between each telemetry channel of the satellite constellation. By sharing parameters and representations of different tasks, obtain the characteristics of each telemetry channel.

[0027] Step 3.2: Use a multi-encoder approach to perform detailed modeling of the characteristics of each telemetry channel in the satellite constellation, and construct a satellite constellation telemetry data prediction model;

[0028] This invention employs a shared feature decoding method with channel decoupling to extract the information most beneficial to the prediction of data from each telemetry channel from the shared features, thus ensuring the accuracy of the data prediction information for each telemetry channel.

[0029] The expression representing the shared information between telemetry channels is:

[0030]

[0031] In formula (10), f(·) is a single encoder model that represents the shared information between channels. To train the telemetry data from time i (inclusive) to time w+i (exclusive), This represents the shared characteristics of each telemetry channel at time i.

[0032] The expression for detailed modeling of the characteristics of each telemetry channel in the satellite constellation is as follows:

[0033]

[0034] In formula (11), f c (·) represents the shared feature decoding model for the c-th telemetry channel. The shared features of each telemetry channel at time i are... Let be the training prediction value for the c-th telemetry channel at the i-th time step.

[0035] Optionally, step 3.1, which involves characterizing the shared information between each telemetry channel of the satellite constellation, includes:

[0036] Step 3.1.1: If the model parameters of the current satellite constellation telemetry data prediction model contain shared information between all telemetry channels, select adaptive parameter learning to represent the relationship between all telemetry channels as the network parameters of the multi-channel shared layer;

[0037] Step 3.1.2: If the model parameters of the current satellite constellation telemetry data prediction model do not contain shared information between all telemetry channels, select training data similarity learning to obtain the similarity of training telemetry data between each telemetry channel, and obtain the representation relationship between each telemetry channel based on the similarity of training telemetry data between each telemetry channel.

[0038] Optionally, step 4, which involves training the satellite constellation telemetry data prediction model, includes:

[0039] Step 4.1: Input the training telemetry data into the satellite constellation telemetry data prediction model to obtain the training prediction values ​​of all telemetry channels at the i-th time step;

[0040] Step 4.2: Based on the training prediction values ​​of all telemetry channels at the i-th time step, obtain the real training telemetry data of the c-th telemetry channel and the training prediction telemetry data of the c-th telemetry channel within the time span w to T1;

[0041] Step 4.3: Calculate the current training variance score ratio of the c-th telemetry channel based on the real training telemetry data and the training prediction telemetry data of the c-th telemetry channel;

[0042] Step 4.4: Based on the current training variance score ratio of the c-th telemetry channel, obtain the proportion of the c-th telemetry channel to the total variance score ratio of all channels in the current training stage;

[0043] Step 4.5: Construct a loss function based on the ratio of the c-th telemetry channel to the variance score of all channels, and obtain the training loss for the current training stage;

[0044] Step 4.6: Adjust the current training phase based on the training loss of the current training phase;

[0045] Step 4.7: Repeat steps 4.3-4.6 until the training loss is less than the preset value, and obtain the trained satellite constellation telemetry data prediction model;

[0046] The formula for calculating the training prediction values ​​of all telemetry channels at the i-th time step is:

[0047]

[0048] In formula (12), Let be the training prediction values ​​for all telemetry channels at the i-th time step. Let be the training prediction value of the c-th telemetry channel at the i-th time step;

[0049] The expressions for the actual training telemetry data and the training predicted telemetry data of the c-th telemetry channel within w~T1 are as follows:

[0050]

[0051] In formulas (13) and (14), For the training and prediction telemetry data of the c-th telemetry channel predicted by the model within a time span of w to T1, y c The actual training telemetry data for the c-th telemetry channel within a time span of w to T1;

[0052] The formula for calculating the current training variance score ratio of the c-th telemetry channel is:

[0053]

[0054] In formula (15), β c It is the current training variance score ratio of the c-th telemetry channel;

[0055] The formula for calculating the proportion of the c-th telemetry channel to the total variance score of all channels in the current training phase is as follows:

[0056]

[0057] In formula (16), λ c It is the proportion of the c-th telemetry channel to the total variance score of all channels in the current training phase, where C is the total number of telemetry channels to be predicted;

[0058] The expression for the loss function is:

[0059]

[0060] In formula (17), It is the real data at the t-th time step of the c-th telemetry channel in the training telemetry data. is the predicted data at the t-th time step of the c-th telemetry channel of the training telemetry data, w is the window length of the data segmented in the preprocessing stage, and T1 is the time span of the training telemetry data.

[0061] Optionally, step 5, which involves obtaining the predicted value for each telemetry channel, includes:

[0062] Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model to obtain the predicted value for each telemetry channel;

[0063] The formula for calculating the telemetry channel prediction value is:

[0064]

[0065] In formula (18), f(·) is a single encoder model that represents the shared information between channels. c (·) represents the shared feature decoding model for the c-th telemetry channel. This is the test prediction value for the c-th telemetry channel at the j-th time step.

[0066] The beneficial effects of this invention are:

[0067] 1. This invention employs a method that combines shared feature extraction with parallel prediction of various telemetry channels for different types of satellites in a satellite constellation to fuse multidimensional parameter information. This allows the models of each telemetry channel to complement and enhance each other during training, thereby improving the overall data monitoring effect.

[0068] 2. This invention employs a channel decoupling shared feature decoding method to extract the information most beneficial to the prediction of data from each telemetry channel from the shared features.

[0069] 3. This invention uses a variance-based scoring method to characterize the learning degree of each channel monitoring task of different satellite models. It adaptively optimizes the model's learning degree for each telemetry channel monitoring task of different satellite models in the satellite constellation, slows down the learning degree of channels that are performing well at the current stage, and speeds up the learning degree of channels that are performing poorly at the current stage. This avoids performance degradation on some difficult telemetry channels due to excessive bias towards a certain easier telemetry channel, and improves the overall performance of the model on all telemetry channels in the satellite constellation. Attached Figure Description

[0070] Figure 1 A flowchart of an adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method provided by the present invention;

[0071] Figure 2 A mathematical modeling block diagram of the satellite constellation telemetry data prediction model provided by this invention;

[0072] Figure 3 This is a diagram showing the channel telemetry data monitoring results based on adaptive parameter learning provided by the present invention. Figure 3 In the diagram, (a) is the voltage of a single battery cell of a satellite in the constellation, (b) is the temperature of the back of the outer panel of the solar cell of a satellite in the constellation, (c) is the temperature of a single cell of structural block a of a satellite in the constellation, (d) is the temperature of a single cell of structural block f of a satellite in the constellation, (e) is the voltage of circuit A of the discharge regulator of a satellite in the constellation, (f) is the voltage of circuit B of the discharge regulator of a satellite in the constellation, (g) is the current of the solar cell power supply array A7 of a satellite in the constellation, and (h) is the charging control status of circuit B of a satellite in the constellation.

[0073] Figure 4 This is a diagram showing the channel telemetry data monitoring results provided by the present invention based on training data similarity learning. Figure 4 In the diagram, (a) represents the voltage of a single battery cell of a satellite in the constellation, (b) represents the discharge switch status of a satellite in the constellation, (c) represents the temperature of the outer panel of the solar cell array of a satellite in the constellation, (d) represents the temperature of the back of the outer panel of the solar cell of a satellite in the constellation, (e) represents the voltage of circuit A2 of the discharge regulator of a satellite in the constellation, (f) represents the voltage of circuit A5 of the discharge regulator of a satellite in the constellation, (g) represents the current of the solar cell power supply array A1 of a satellite in the constellation, and (h) represents the current of the solar cell power supply array A7 of a satellite in the constellation. Example

[0074] Combination Figure 1-4 This embodiment will be described as follows: Figure 1 As shown in this embodiment, an adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method includes:

[0075] S1: Acquire time series data of satellite constellation telemetry;

[0076] S2: Preprocessing of satellite constellation telemetry time series data;

[0077] S201: Divide the satellite constellation telemetry time series data into training telemetry datasets and test telemetry datasets;

[0078] S20101: The ratio of the training telemetry dataset to the test telemetry dataset is 8:2. The training telemetry dataset and the test telemetry dataset are as follows:

[0079]

[0080] In formulas (1)-(8), For the training telemetry data of telemetry channel c at time t1, For the test telemetry data of telemetry channel c at time t2, X train This refers to the training telemetry data after being segmented into windows of length w, where w ranges from 15 to 35. To train the telemetry data from time i (inclusive) to time w+i (exclusive), Y train For the corresponding X train In each window, the actual training telemetry data is predicted p time steps backward, where p ranges from 1 to 20. To train the telemetry data from time w+i (inclusive) to time w+p+i (exclusive), X test The data was segmented into windows of length w for testing telemetry data. To test the data values ​​of Y from time j (inclusive) to time w+j (exclusive) in the telemetry data, test For the corresponding X test In each window, the actual training telemetry data is predicted p time steps backward. To test the data values ​​from time w+j (inclusive) to time w+p+j (exclusive) in the telemetry data, T1 is the time span of the training telemetry data, ranging from 200 to 20000, and T2 is the time span of the test telemetry data, ranging from 200 to 20000.

[0081] S202: Standardize the training telemetry dataset by channel;

[0082] The expression for standardization is:

[0083]

[0084] In formula (9), For the training telemetry data of telemetry channel c at time t1, μ(X)trainc ) represents the mean of all training telemetry data for telemetry channel c, σ(X) trainc ) represents the standard deviation of all training telemetry data for telemetry channel c.

[0085] S3: Building a satellite constellation telemetry data prediction model;

[0086] like Figure 2 As shown, the satellite constellation telemetry data prediction model consists of two parts. The first is a multi-telemetry channel sharing layer, which uses a front-end single encoder structure to represent the shared information between channels. By sharing parameters and representations of different tasks, it can effectively capture the correlation and commonalities between tasks. This sharing mechanism helps the model learn more general features for each task, thereby improving the overall generalization ability and reducing overfitting. The second is a task-specific layer, which uses a multi-decoder approach to finely model the characteristics of different telemetry channels. At the same time, in order to balance the learning degree of each telemetry channel and avoid the model from overfitting to simple tasks and underfitting to complex tasks, an adaptive learning rate adjustment scheme based on the ratio of prediction error variance to original model variance is adopted. This slows down the training speed of telemetry channels with better monitoring performance in the current training stage, while strengthening the training of telemetry channels with poor monitoring performance in the current stage, so as to achieve higher multi-dimensional telemetry channel data monitoring performance.

[0087] S301: A single encoder is used to represent the shared information between the telemetry channels of the satellite constellation. By sharing the parameters and representations of different tasks, the correlation and commonalities between tasks can be effectively captured. This sharing mechanism helps the model learn more general features on each task, thereby improving the overall generalization ability and reducing overfitting.

[0088] The expression representing the shared information between telemetry channels is:

[0089]

[0090] In formula (10), f(·) is a single encoder model that represents the shared information between channels. To train the telemetry data from time i (inclusive) to time w+i (exclusive), This represents the shared characteristics of each telemetry channel at time i.

[0091] S30101: There are two modes for representing the relationship between channels. One is adaptive parameter learning, which refers to representing the relationship as the parameters of the network.

[0092] When the shared parameter learning method among channels is adaptive parameter learning, the telemetry data monitoring results of each channel in the satellite constellation are as follows: Figure 3As shown, the mean absolute error across all monitoring channels is 0.0152, the root mean square error is 0.0223, and the R-squared score is 0.9232.

[0093] S30102: The second is learning based on the similarity of training data, which means using the similarity of historical data as a reference. The greater the similarity, the greater the correlation.

[0094] When the shared parameters between channels are learned based on the similarity of training data, the telemetry data monitoring results of each channel in the satellite constellation are as follows: Figure 4 As shown, the mean absolute error across all monitoring channels is 0.0261, the root mean square error is 0.0409, and the R-squared fraction is 0.8923, indicating that the present invention can achieve high-precision multi-channel parallel satellite constellation telemetry data monitoring.

[0095] S302: Employs a multi-decoder approach to perform detailed modeling of the characteristics of different telemetry channels in the satellite constellation;

[0096] The expression for detailed modeling of the characteristics of each telemetry channel in the satellite constellation is as follows:

[0097]

[0098] In formula (11), f c (·) represents the shared feature decoding model for the c-th telemetry channel. The shared features of each telemetry channel at time i are... Let be the training prediction value for the c-th telemetry channel at the i-th time step.

[0099] S4: Training of satellite constellation telemetry data prediction model;

[0100] Predictive modeling for telemetry channels across different satellite constellations is a complex and heterogeneous task. A real satellite constellation may consist of multiple satellites carrying different payloads, with varying types of telemetry parameters. Furthermore, telemetry data typically includes multiple channels, each corresponding to different sensor or system parameters, such as temperature, voltage, attitude, and position. Because the telemetry parameters of different channels exhibit different temporal variations due to the different characteristics of their corresponding physical components, the model's learning ability, speed, and efficiency differ for different data types. To balance the learning intensity of different satellite models and telemetry channels for prediction tasks, and to avoid overfitting to simple tasks and underfitting to complex tasks, it is necessary to slow down the training speed of telemetry channels with better monitoring performance in the current training phase, while simultaneously strengthening the training of telemetry channels with poor monitoring performance in the current phase. The ratio of the variance in the prediction error to the variance of the original data can reflect the predictive performance of the model during the current training phase to some extent. A larger ratio means that the model's prediction results contain more noise and have not yet fully learned the temporal variation patterns of the current telemetry channel. Conversely, the closer this ratio is to 0, the closer the prediction error is to 0, indicating better monitoring performance of the temporal evolution patterns of the current telemetry channel data. Therefore, this method adopts an adaptive learning rate adjustment scheme based on the ratio of the prediction error variance to the original model variance to achieve higher monitoring performance of multidimensional telemetry channel data.

[0101] S401: Input the training telemetry data into the satellite constellation telemetry data prediction model to obtain the training prediction values ​​of all telemetry channels at the i-th time step;

[0102] S402: Based on the training prediction values ​​of all telemetry channels at the i-th time step, obtain the real training telemetry data of the c-th telemetry channel and the training prediction telemetry data of the c-th telemetry channel within the time span w to T1.

[0103] S403: The current training variance score ratio of the c-th telemetry channel is calculated based on the real training telemetry data and the training prediction telemetry data of the c-th telemetry channel.

[0104] S404: Based on the current training variance score ratio of the c-th telemetry channel, obtain the proportion of the c-th telemetry channel to the total variance score ratio of all channels in the current training stage;

[0105] S405: Construct a loss function based on the ratio of the c-th telemetry channel to the variance score of all channels, and obtain the training loss for the current training stage;

[0106] S406: Adjust the current training phase based on the training loss of the current training phase;

[0107] S407: Repeat S403-S406 until the training loss is less than the preset value, and obtain the trained satellite constellation telemetry data prediction model;

[0108] The formula for calculating the training prediction values ​​of all telemetry channels at the i-th time step is:

[0109]

[0110] In formula (12), Let be the training prediction values ​​for all telemetry channels at the i-th time step. Let be the training prediction value of the c-th telemetry channel at the i-th time step;

[0111] The expressions for the actual training telemetry data and the training predicted telemetry data of the c-th telemetry channel within w~T1 are as follows:

[0112]

[0113] In formulas (13) and (14), For the training and prediction telemetry data of the c-th telemetry channel predicted by the model within a time span of w to T1, y c The actual training telemetry data for the c-th telemetry channel within a time span of w to T1;

[0114] The formula for calculating the current training variance score ratio of the c-th telemetry channel is:

[0115]

[0116] In formula (15), β c It is the current training variance score ratio of the c-th telemetry channel;

[0117] The formula for calculating the proportion of the c-th telemetry channel to the total variance score of all channels in the current training phase is as follows:

[0118]

[0119] In formula (16), λ c It is the proportion of the c-th telemetry channel to the total variance score of all channels in the current training phase, where C is the total number of telemetry channels to be predicted;

[0120] The expression for the loss function is:

[0121]

[0122] In formula (17), It is the real data at the t-th time step of the c-th telemetry channel in the training telemetry data. is the predicted data at the t-th time step of the c-th telemetry channel of the training telemetry data, w is the window length of the data segmented in the preprocessing stage, and T1 is the time span of the training telemetry data.

[0123] S5: Multi-channel prediction based on satellite constellation telemetry data prediction model, and real-time monitoring of satellite constellation telemetry data;

[0124] Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model to obtain the predicted value for each telemetry channel;

[0125] The formula for calculating the telemetry channel prediction value is:

[0126]

[0127] In formula (18), f(·) is a single encoder model that represents the shared information between channels. c (·) represents the shared feature decoding model for the c-th telemetry channel. This is the test prediction value for the c-th telemetry channel at the j-th time step.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. An adaptive, multi-task balanced, parallel satellite constellation telemetry data monitoring method, characterized in that, The steps of the adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method include: Step 1: Acquire satellite constellation telemetry time series data; Step 2: Preprocess the telemetry time series data of the satellite constellation; Step 3: Construct a satellite constellation telemetry data prediction model; Step 3, which involves constructing a satellite constellation telemetry data prediction model, includes the following steps: Step 3.1: Use a single encoder to characterize the shared information between each telemetry channel of the satellite constellation. By sharing parameters and representations of different tasks, obtain the characteristics of each telemetry channel. Step 3.1, which involves characterizing the shared information between each telemetry channel of the satellite constellation, includes: Step 3.1.1: If the model parameters of the current satellite constellation telemetry data prediction model contain shared information between all telemetry channels, select adaptive parameter learning to represent the relationship between all telemetry channels as network parameters of a multi-channel shared layer; Step 3.1.2: If the model parameters of the current satellite constellation telemetry data prediction model do not contain the shared information between all telemetry channels, select training data similarity learning to obtain the similarity of training telemetry data between each telemetry channel, and obtain the representation relationship between each telemetry channel based on the similarity of training telemetry data between each telemetry channel. Step 3.2: Use a multi-encoder approach to perform detailed modeling of the characteristics of each telemetry channel in the satellite constellation, and construct a satellite constellation telemetry data prediction model; The expression representing the shared information between telemetry channels is: （10） In formula (10), A single encoder model for representing shared information between channels. To train time in telemetry data Time Data values, including time Including time points ,time Not including time points , For the first Shared characteristics of various telemetry channels at different times; The expression for detailed modeling of the characteristics of each telemetry channel in the satellite constellation is as follows: （11） In formula (11), For the first A shared feature decoding model for each telemetry channel. For the first Shared characteristics of each telemetry channel at a given time. For the first The first time step Training prediction values ​​for each telemetry channel; Step 4: Train the satellite constellation telemetry data prediction model based on the preprocessed satellite constellation telemetry time series data; Step 4, which involves training the satellite constellation telemetry data prediction model, includes the following steps: Step 4.1: Input the trained telemetry data into the satellite constellation telemetry data prediction model to obtain the first... i Training predictions for all telemetry channels at each time step; Step 4.2: Based on the aforementioned... i The training prediction values ​​for all telemetry channels at each time step are obtained over a time span of [time span]. The first Real training telemetry data for the 3D telemetry channel and the 1D telemetry data. Training and predicting telemetry data for multiple telemetry channels; Step 4.3: Based on the aforementioned... Real training telemetry data for the 3D telemetry channel and the 1D telemetry data. The training prediction telemetry data of the telemetry channel is used to calculate the first telemetry data. The current training variance score ratio of the telemetry channel; Step 4.4: Based on the aforementioned... The current training variance score ratio of the telemetry channel is used to obtain the current training phase. The proportion of remote sensing channels in the overall variance score of all channels; Step 4.5: Based on the aforementioned... The loss function is constructed by taking the proportion of the variance scores of the telemetry channels to the total variance scores of all channels, and the training loss for the current training stage is obtained. Step 4.6: Adjust the current training stage based on the training loss of the current training stage; Step 4.7: Repeat steps 4.3-4.6 until the training loss is less than the preset value, and obtain the trained satellite constellation telemetry data prediction model; The formula for calculating the training prediction values ​​of all telemetry channels at the i-th time step is: （12） In formula (12), For the first i The training prediction values ​​of all telemetry channels at each time step. For the first The first time step Training prediction values ​​for each telemetry channel; The first Real training telemetry data for the 3D telemetry channel and the 1D telemetry data. The expressions for training and predicting telemetry data for the 3D telemetry channels are as follows: （13） （14） In formulas (13) and (14), For a time span of The model predicts the first Training and predicting telemetry data for 3D telemetry channels. For a time span of The first Real training telemetry data for the telemetry channel; No. The formula for calculating the current training variance score ratio of the telemetry channel is: （15） In formula (15), It is the first The current training variance score ratio of the telemetry channel; Current training phase The formula for calculating the proportion of telemetry channels in the overall variance score of all channels is as follows: （16） In formula (16), This is the current training phase. The proportion of remote sensing channels in the overall variance score of all channels. This represents the total number of telemetry channels to be predicted. The expression for the loss function is: （17） In formula (17), It is the training telemetry data. The telemetry channel number Real data at each time step It is the training telemetry data. The telemetry channel number Predicted data at each time step It is the window length for splitting data during the preprocessing stage. It is the time span of the training telemetry data; Step 5: Obtain the predicted value of each telemetry channel based on the trained satellite constellation telemetry data prediction model, and monitor the satellite constellation telemetry data in real time.

2. The adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method according to claim 1, characterized in that, Step 2, which involves preprocessing the satellite constellation telemetry time series data, includes: Step 2.1: Divide the satellite constellation telemetry time series data into a training telemetry dataset and a test telemetry dataset, with the ratio of the training set to the test set being 8:2; Step 2.2: Standardize the training telemetry dataset by channel; The expression for standardization is: （1） In formula (1), For telemetry channel exist Training telemetry data at any given time, For telemetry channel The mean of all training telemetry data For telemetry channel The standard deviation of all training telemetry data.

3. The adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method according to claim 2, characterized in that, The training telemetry dataset and test telemetry dataset after partitioning in step 2.1 include , , , , , , and ; The expressions for the training telemetry dataset and the test telemetry dataset are as follows: （2） （3） （4） （5） （6） （7） （8） （9） In formulas (2)-(9), For telemetry channel exist Training telemetry data at any given time, For telemetry channel exist Test telemetry data at any time, To perform windowing on the training telemetry data with a length of [missing information] The segmented data, window length The range is 15-35. To train time in telemetry data Time Data values, time Including time points , Not including time points ; For the corresponding Forward prediction of each window Real training telemetry data at each time step The range is 1-20. To train time in telemetry data Time Data values, time Including time points ,time Not including time points ; To test telemetry data, a window with a length of [missing information] is used. The data after segmentation, To test time in telemetry data Time Data values, time Including time points ,time Not including time points ; For the corresponding Forward prediction of each window Real training telemetry data at each time step To test time in telemetry data Time Data values, time Including time points ,time Not including time points ; To train the time span of telemetry data, The range is 200-20000. To test the time span of telemetry data, The range is 200-20000.

4. The adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method according to claim 1, characterized in that, The satellite constellation telemetry data prediction model in step 3 includes a multi-channel shared layer and a mission-specific layer; The multi-channel sharing layer is used to characterize the shared information between each telemetry channel and capture the correlation and commonalities between different tasks; The specific mission layer is used to perform fine-grained modeling of the characteristics of each telemetry channel in the satellite constellation.

5. The adaptive multi-task balanced parallel satellite constellation telemetry data monitoring method according to claim 2, characterized in that, Step 5, which involves obtaining the predicted value for each telemetry channel, includes: The test telemetry dataset is input into the trained satellite constellation telemetry data prediction model to obtain the predicted value for each telemetry channel; The formula for calculating the telemetry channel prediction value is: （18） In formula (18), A single encoder model for representing shared information between channels. For the first A shared feature decoding model for each telemetry channel. For the first The first time step Test prediction values ​​for each telemetry channel.

Images