Satellite constellation telemetry anomaly detection method based on adaptive spatio-temporal feature fusion

Abstract

The application provides a satellite constellation telemetry anomaly detection method based on adaptive space-time feature fusion, comprising the following steps: step 1: dividing historical satellite constellation telemetry time series data of a constellation network into a training telemetry data set and a test telemetry data set; step 2: preprocessing the training telemetry data set; step 3: constructing a satellite constellation telemetry data prediction model; step 4: training the satellite constellation telemetry data prediction model based on the training telemetry data set; step 5: inputting the test telemetry data set into the trained satellite constellation telemetry data prediction model to obtain a satellite constellation telemetry data prediction result, and performing real-time anomaly detection on the satellite constellation telemetry data prediction. Through the alternate feature extraction on the time scale and the space scale, the application can output a prediction result more inclined to a normal operation state, and is beneficial to realize higher-precision anomaly detection.

Description

Technical Field

[0001] This invention relates to an adaptive spatiotemporal feature fusion method for satellite constellation telemetry anomaly detection, belonging to the field of satellite anomaly detection technology. Background Technology

[0002] In practical applications, satellites may experience anomalies due to various reasons during operation, such as equipment failure, changes in the external environment, or operational errors. If these anomalies are not detected and handled in a timely manner, they may lead to a decline in satellite performance or even major malfunctions, affecting the smooth progress of the mission. Therefore, how to effectively detect and warn of anomalies has become a key issue in satellite management. With the increasing maturity and sophistication of microsatellite manufacturing and launch, their costs have been greatly reduced. Complex missions can be accomplished collaboratively by a constellation of satellites responsible for different sub-tasks. Because satellites in a constellation network need to work together to complete a specific task, there is a certain degree of correlation between the various telemetry channels of different satellite models. In addition, the changes in the satellite's own operating status and the high coupling between internal systems mean that the future changes in telemetry data are not only related to its historical temporal evolution trend, but also closely related to the current operating status of other telemetry parameters with similar operating modes. Therefore, capturing the correlation information between various telemetry channels, combining spatial characteristics with the temporal evolution characteristics of the telemetry data itself, and utilizing the normal data when an anomaly occurs in a certain telemetry channel in its early stage before it spreads to all its related systems, to strengthen the construction of a prediction model for the normal working mode of each telemetry channel, and thus achieve a more accurate anomaly detection method, has become a research hotspot.

[0003] In existing technologies, the National University of Defense Technology of the Chinese People's Liberation Army (CN107644148) uses the maximum mutual information coefficient to obtain the spatial correlation of telemetry parameter values. It calculates the mutual information coefficient based on the scatter plot of historical telemetry parameter data corresponding to the satellite in normal and abnormal states, and determines the distribution of the mutual information coefficient to judge whether there is anomaly in the multi-parameter correlation state. However, it does not take into account that in actual scenarios, large-scale satellite constellations are still in the early stages of development, and the number of abnormal samples in the reserve telemetry data is scarce. Normal samples account for the majority of the telemetry data, and in some scenarios, it is difficult to establish a relatively difficult-to-accurately locate the interval where the anomaly occurs using the limited abnormal samples. Meanwhile, the time and computational costs of this distance-based anomaly detection method increase with the dimensionality of the telemetry data and the selected time window. With the development of aerospace technology, spacecraft telemetry data resources are becoming increasingly abundant. Sometimes, the computing resources of the computing platform are difficult to achieve real-time computing effects under such high-dimensional data input. The technical solution of this invention only uses historically stored normal telemetry data and adopts a graph network with an importance sampling strategy to adaptively capture the relationship between various telemetry channels. It can not only adapt to the changing mission scenarios, but also greatly reduce the computational complexity through relation parameter matrix decomposition. Changguang Satellite Technology Co., Ltd. (CN110703738) proposed a supervised local linear embedding fault detection method for satellite attitude control systems. It acquires high-dimensional raw telemetry satellite data, performs feature analysis and preprocessing on the acquired raw satellite telemetry data, uses the SLLE algorithm to reduce the dimensionality of the preprocessed raw satellite telemetry data, obtains low-dimensional embedding feature information of the satellite control system telemetry data, and uses SPE statistics to complete fault detection. However, this approach cannot perform fine-grained anomaly localization at the component level and is limited to satellite control systems, making it difficult to apply to large-scale satellite constellations. It also struggles to capture both long-term trends (the working patterns of different satellites, determined by their designed operating modes) and short-term fluctuations (fluctuations of varying degrees caused by the operational space environment at different times) in telemetry data. In contrast, the technical solution of this invention extracts feature networks through channel coupling relationships to characterize the mutual influence relationships between telemetry data at the component level without dimensionality reduction, and models the features related to all components, which is beneficial for achieving more precise anomaly component localization. Summary of the Invention

[0004] This invention addresses the technical problem in existing satellite data monitoring technologies that struggle to simultaneously capture both long-term trends and short-term fluctuations in telemetry data. It proposes an adaptive spatiotemporal feature fusion method for detecting telemetry anomalies in satellite constellations.

[0005] The technical solution adopted by this invention to solve the above problems is as follows: This invention proposes an adaptive spatiotemporal feature fusion method for satellite constellation telemetry anomaly detection, comprising:

[0006] Step 1: Divide the historical satellite constellation telemetry time series data of the constellation network into training telemetry datasets and test telemetry datasets;

[0007] Step 2: Preprocess the training telemetry dataset;

[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 training telemetry dataset;

[0010] Step 5: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model to obtain the satellite constellation telemetry data prediction results, and perform real-time anomaly detection on the satellite constellation telemetry data prediction.

[0011] Optionally, in step 1, the ratio of the training telemetry dataset to the test telemetry dataset is 8:2;

[0012] The expressions for the training telemetry dataset and the test telemetry dataset are as follows:

[0013]

[0014]

[0015] 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.

[0016] Optionally, the preprocessing of the training telemetry dataset in step 2 includes:

[0017] The training telemetry dataset is standardized to obtain a preprocessed training telemetry dataset.

[0018] The expression for standardizing the training telemetry dataset is:

[0019]

[0020] 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.

[0021] Optionally, the satellite constellation telemetry data prediction model in step 3 includes a temporal feature extraction module and a spatial feature extraction block;

[0022] The time feature extraction module is used to extract multi-level dependency information of all telemetry channels of the satellite constellation on historical satellite constellation telemetry time series data;

[0023] The spatial feature extraction module is used to model all telemetry channels of the satellite constellation and to fit and sort the adjacent nodes in the telemetry channels.

[0024] Optionally, step 3, which involves constructing a satellite constellation telemetry data prediction model, includes:

[0025] Step 3.1: Based on the time feature extraction module, analyze the low-frequency and high-frequency components in the historical satellite constellation telemetry time series data respectively, and extract the multi-level dependency information of all telemetry channels of the satellite constellation on the historical satellite constellation telemetry time series data;

[0026] This invention employs a high-frequency and low-frequency decomposition method to fit and capture the dependence of various telemetry time series of a satellite constellation on historical data at different frequency distribution ranges, thereby analyzing the multi-level change patterns of the satellite and achieving precise component-level localization of anomalies.

[0027] Step 3.2: Based on the spatial feature extraction module, adaptively learn the adjacency matrix between all telemetry channels of the satellite constellation, determine the coupling relationship between all telemetry channels and model it, and obtain the adjacent nodes with related relationships based on the adjacency matrix, and fit and sort the current node according to the importance of the adjacent nodes with related relationships.

[0028] Step 3.3: Extract information from historical satellite constellation telemetry time series data alternately through the time feature extraction module and the spatial feature extraction module to obtain the historical telemetry sequences of all telemetry channels in the input satellite constellation telemetry data prediction model, and output the prediction mapping function f(·) for all prediction time window telemetry data;

[0029] This invention improves the detection accuracy of satellite constellation telemetry data by extracting features alternately on both temporal and spatial scales through the satellite constellation telemetry data.

[0030] The expressions for the low-frequency and high-frequency components are as follows:

[0031]

[0032] In formulas (10) and (11), Let be the low-frequency component of channel c at time t. Let be the high-frequency component of channel c at time t. Let AvgPool(·) represent the average pooling for the telemetry data of channel c at time t.

[0033] The expression for the adjacency matrix among all telemetry channels is:

[0034] M = sigmoid(EE) T (12)

[0035] In formula (12), E represents the learning parameters of the satellite constellation telemetry data prediction model. T Transpose itself;

[0036] The expression for determining the association relationship between adjacent nodes is:

[0037]

[0038] In formula (13), A ij For each element in the adjacency matrix, A ij A value of 1 indicates that there is a relationship between the two nodes, A ij A value of 0 indicates that the two nodes are not related.

[0039] The expression for fitting and sorting adjacent nodes is:

[0040]

[0041] α ij =softmax(α) i,· (15)

[0042] In formulas (14) and (15), α ij Let A represent the importance of node j to node i. Concat(·) concatenates two tensors along a common dimension. ij For each element in the adjacency matrix of the telemetry channel nodes, softmax(·) is the normalized exponential function;

[0043] The expression for the historical telemetry sequence is:

[0044]

[0045] In formulas (16) and (17), The predicted telemetry time series data corresponding to the training dataset from time point w+i (inclusive) to w+p+i (exclusive). The predicted telemetry time series data corresponding to the test dataset from time point w+j (inclusive) to w+p+j (exclusive). To train the telemetry data from time i (inclusive) to time w+i (exclusive), To test the data values ​​from time j (inclusive) to time w+j (exclusive) in the telemetry data.

[0046] Optionally, step 4, which involves training the satellite constellation telemetry data prediction model, includes:

[0047] Step 4.1: Input the training telemetry dataset and obtain the corresponding predicted telemetry time series data;

[0048] Step 4.2: Calculate the root mean square error between the predicted telemetry time series data and the actual telemetry time series data corresponding to the training telemetry dataset;

[0049] Step 4.3: Adjust the current training phase based on the root mean square error of the current training phase;

[0050] Step 4.4: Repeat steps 4.1-4.3 until the root mean square error is less than the preset value, and obtain the trained satellite constellation telemetry data prediction model;

[0051] The expression for the root mean square error loss function is:

[0052]

[0053] In formula (18), This refers to the actual telemetry time series data corresponding to the training telemetry dataset from time point w+i (inclusive) to w+p+i (exclusive). T1 represents the predicted telemetry time series data corresponding to the training telemetry dataset from time point w+i (inclusive) to w+p+i (exclusive), where w is the window length of the data segmented in the preprocessing stage, and T1 is the time span of the training telemetry data.

[0054] Optionally, step 5, which involves real-time anomaly detection of the predicted satellite constellation telemetry data, includes:

[0055] Step 5.1: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model and calculate the absolute prediction error at each time step;

[0056] Step 5.2: Calculate the smoothed prediction error of the test telemetry data using the exponentially weighted moving average method and the absolute prediction error at each time step;

[0057] Step 5.3: Calculate the detection threshold for all telemetry channels based on the smoothed prediction error of the test telemetry data;

[0058] Step 5.4: Perform anomaly detection on subsequent time series based on the detection thresholds of all telemetry channels. If the smoothing prediction error at a time step is greater than the detection threshold, it is considered an anomaly.

[0059] The formula for calculating absolute prediction error is:

[0060]

[0061] In formulas (19) and (20), To test the actual data at the j-th time step of the telemetry data, To test the predicted data at the j-th time step of the telemetry data, e (j) To test the absolute prediction error of the telemetry data at the j-th time step, e is the absolute prediction error of the test telemetry data, w is the window length of the data segmented in the preprocessing stage, and T2 is the time span of the test telemetry data.

[0062] The formula for calculating the smoothing prediction error is:

[0063]

[0064] In formulas (21)-(24), span is the window length of the weighted average algorithm, α is the smoothing parameter satisfying 0<α≤1, which determines the weight of the most recent observation, and e s (j) To test the smoothing prediction error of telemetry data at the j-th time step, es To test the smoothing prediction error of telemetry data;

[0065] The formula for calculating the detection threshold of all telemetry channels is:

[0066] h=δ·T2(25)

[0067]

[0068] In formulas (25) and (26), h is the length of the early anomaly-free window used to determine the anomaly detection threshold, δ is the percentage of the early anomaly-free window in the total time span of the test data, z is a constant between 1 and 3.5, which is a vector and can be selected for different channels, T2 is the time span of the test telemetry data, and ò is the set anomaly detection threshold. σ is the smoothed prediction error value from 0 (inclusive) to h (exclusive), μ(·) is the mean of the data within the brackets, and σ(·) is the standard deviation of the data within the brackets.

[0069] The beneficial effects of this invention are:

[0070] 1. This invention uses the ideas of adaptive relationship learning and neighbor channel importance sampling to model the collaborative relationship between satellites in a satellite constellation and the internal coupling relationship of telemetry parameters.

[0071] 2. This invention uses a high-frequency and low-frequency decomposition method to fit and capture the dependence of each telemetry time series of a satellite constellation on historical data at different frequency distribution ranges in order to analyze the multi-level change patterns of the satellite.

[0072] 3. By extracting features alternately on time and space scales, this invention can output prediction results that are more biased towards normal operation, which is beneficial for achieving higher accuracy in anomaly detection. Attached Figure Description

[0073] Figure 1 A flowchart of a satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion provided by the present invention;

[0074] Figure 2 A mathematical modeling block diagram for satellite constellation telemetry anomaly detection provided by this invention;

[0075] Figure 3 This is an anomaly detection result diagram of the telemetry time series under the first set of parameter settings provided by the present invention. Figure 3In the following, (a) the voltage of a single battery cell of a satellite in the constellation, (b) the discharge current of battery group A of a satellite in the constellation, (c) the discharge current of battery group B of a satellite in the constellation, (d) the output current of the discharge regulator of channel B of a satellite in the constellation, (e) the temperature of the shunt of a satellite in the constellation, and (f) the status of the charging controller of channel A of a satellite in the constellation.

[0076] Figure 4 This is an image showing the anomaly detection results of the telemetry time series under the second set of parameter settings provided by the present invention. Figure 4 In the following, (a) the voltage of a single battery cell of a satellite in the constellation, (b) the discharge current of battery group A of a satellite in the constellation, (c) the discharge current of battery group B of a satellite in the constellation, (d) the output current of the discharge regulator of channel B of a satellite in the constellation, (e) the temperature of the shunt of a satellite in the constellation, and (f) the status of the charging controller of channel A of a satellite in the constellation. Example

[0077] Example 1

[0078] Combination Figure 1-2 This embodiment will be described as follows: Figure 1 As shown in this embodiment, a satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion includes:

[0079] S1: Divide the historical satellite constellation telemetry time series data of the constellation network into training telemetry datasets and test telemetry datasets;

[0080] S101: The ratio of the training telemetry dataset to the test telemetry dataset is 8:2;

[0081] The expressions for the training telemetry dataset and the test telemetry dataset are as follows:

[0082]

[0083] 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), Xtest 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.

[0084] S2: Preprocess the training telemetry dataset;

[0085] S201: Standardize the training telemetry dataset to obtain a preprocessed training telemetry dataset;

[0086] The expression for standardizing the training telemetry dataset is:

[0087]

[0088] 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.

[0089] S3: Construct a satellite constellation telemetry data prediction model;

[0090] like Figure 2 As shown, the satellite constellation telemetry data prediction model includes a temporal feature extraction module and a spatial feature extraction block. The temporal feature extraction module is a channel coupling relationship extraction network implemented using adaptive relationship learning and neighbor channel importance sampling. The spatial feature extraction block is a feature extraction network that uses high-frequency short-term fluctuations and low-frequency long-term trend decomposition to amplify and capture the dependence of telemetry time series on historical data at different frequency distribution ranges. The two feature extraction networks are stacked together. Through the cross-use of the two types of networks, the dependence relationship learning in both time and space can be achieved, improving the model's ability to predict normal patterns in satellite telemetry data and facilitating higher-precision anomaly detection.

[0091] S301: The time feature extraction module is used to analyze and extract the multi-level dependency information of all telemetry channels of the satellite constellation on historical data by analyzing the low-frequency components (evolutionary trends and long-term slow fluctuations) and high-frequency components (short-term fluctuations) in the historical satellite constellation telemetry time series data.

[0092] The expressions for the low-frequency and high-frequency components are as follows:

[0093]

[0094] In formulas (10) and (11), Let be the low-frequency component of channel c at time t. Let be the high-frequency component of channel c at time t. Let AvgPool(·) represent the average pooling for the telemetry data of channel c at time t.

[0095] The expression for the adjacency matrix among all telemetry channels is:

[0096] S302: The spatial feature extraction block uses an adaptive adjacency matrix learning strategy to model whether there is coupling between all channels of the satellite constellation. Based on the adjacency matrix, it obtains the adjacent nodes with related relationships and fits and sorts the current node according to the importance of the adjacent nodes with related relationships, so as to achieve more refined multi-channel information fusion.

[0097] The expression for the adjacency matrix among all telemetry channels is:

[0098] M = sigmoid(EE) T (12)

[0099] In formula (12), E represents the learning parameters of the satellite constellation telemetry data prediction model. T Transpose itself;

[0100] The expression for determining the association relationship between adjacent nodes is:

[0101]

[0102] In formula (13), A ij For each element in the adjacency matrix, A ij A value of 1 indicates that there is a relationship between the two nodes, A ij A value of 0 indicates that the two nodes are not related.

[0103] The expression for fitting and sorting adjacent nodes is:

[0104]

[0105] α ij=softmax(α) i,· (15)

[0106] In formulas (14) and (15), α ij Let A represent the importance of node j to node i. Concat(·) concatenates two tensors along a common dimension. ij For each element in the adjacency matrix of the telemetry channel nodes, softmax(·) is the normalized exponential function;

[0107] S303: By alternately extracting information of historical telemetry parameters through the temporal feature extraction module and the spatial feature extraction block, the model representation of the historical telemetry sequence of each channel of the input satellite constellation can be obtained, and the prediction mapping function f(·) corresponding to the telemetry data of the prediction time window can be output.

[0108] The expression for the historical telemetry sequence is:

[0109]

[0110] In formulas (16) and (17), The predicted telemetry time series data corresponding to the training dataset from time point w+i (inclusive) to w+p+i (exclusive). The predicted telemetry time series data corresponding to the test dataset from time point w+j (inclusive) to w+p+j (exclusive). To train the telemetry data from time i (inclusive) to time w+i (exclusive), To test the data values ​​from time j (inclusive) to time w+j (exclusive) in the telemetry data.

[0111] S4: Train the satellite constellation telemetry data prediction model based on the training telemetry dataset;

[0112] S401: Input the training telemetry dataset and obtain the predicted telemetry time series data corresponding to the training telemetry dataset;

[0113] S402: Calculate the root mean square error between the predicted telemetry time series data and the actual telemetry time series data corresponding to the training telemetry dataset;

[0114] S403: Adjust the current training phase based on the root mean square error of the current training phase;

[0115] S404: Repeat S401-S403 until the root mean square error is less than the preset value to obtain the trained satellite constellation telemetry data prediction model.

[0116] The expression for the root mean square error loss function is:

[0117]

[0118] In formula (18), This refers to the actual telemetry time series data corresponding to the training telemetry dataset from time point w+i (inclusive) to w+p+i (exclusive). T1 represents the predicted telemetry time series data corresponding to the training telemetry dataset from time point w+i (inclusive) to w+p+i (exclusive), where w is the window length of the data segmented in the preprocessing stage, and T1 is the time span of the training telemetry data.

[0119] S5: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model to obtain the satellite constellation telemetry data prediction results, and perform real-time anomaly detection on the satellite constellation telemetry data prediction;

[0120] S501: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model and calculate the absolute prediction error at each time step;

[0121] S502: The smoothed prediction error of the test telemetry data is calculated using the exponentially weighted moving average method and the absolute prediction error at each time step;

[0122] S503: The detection threshold for all telemetry channels is calculated based on the smoothed prediction error of the test telemetry data;

[0123] S504: Perform anomaly detection on subsequent time series based on the detection thresholds of all telemetry channels. If the smoothing prediction error at a time step is greater than the detection threshold, it is considered an anomaly.

[0124] The formula for calculating absolute prediction error is:

[0125]

[0126] In formulas (19) and (20), To test the actual data at the j-th time step of the telemetry data, To test the predicted data at the j-th time step of the telemetry data, e (j) To test the absolute prediction error of the telemetry data at the j-th time step, e is the absolute prediction error of the test telemetry data, w is the window length of the data segmented in the preprocessing stage, and T2 is the time span of the test telemetry data.

[0127] The formula for calculating the smoothing prediction error is:

[0128]

[0129] In formulas (21)-(24), span is the window length of the weighted average algorithm, α is the smoothing parameter satisfying 0<α≤1, which determines the weight of the most recent observation, and e s (j) To test the smoothing prediction error of telemetry data at the j-th time step, e s To test the smoothing prediction error of telemetry data;

[0130] The formula for calculating the detection threshold of all telemetry channels is:

[0131] h=δ·T2(25)

[0132]

[0133] In formulas (25) and (26), h is the length of the early anomaly-free window used to determine the anomaly detection threshold, δ is the percentage of the early anomaly-free window in the total time span of the test data, z is a constant between 1 and 3.5, which is a vector and can be selected for different channels, T2 is the time span of the test telemetry data, and ò is the set anomaly detection threshold. σ is the smoothed prediction error value from 0 (inclusive) to h (exclusive), μ(·) is the mean of the data within the brackets, and σ(·) is the standard deviation of the data within the brackets.

[0134] Example 2

[0135] Combination Figure 3 This embodiment will be described in detail. The first set of parameters in this embodiment includes:

[0136] The time span of the training telemetry data is T1: 12000;

[0137] The time span for testing telemetry data is T2:8000;

[0138] The validation set partition ratio during training is 0.2.

[0139] Batch size during model training: 32;

[0140] Maximum training epochs: 30;

[0141] The time window length for telemetry historical time series slices is w: 25;

[0142] Telemetry prediction time window length p:10;

[0143] The window length for the EWMA weighted average is 70.

[0144] The proportion of early-stage non-abnormal window lengths is δ: 5%;

[0145] z:3.

[0146] The anomaly detection effect of this embodiment on satellite constellation telemetry time series data is as follows: Figure 3 As shown, the anomaly detection precision obtained on actual telemetry data with anomalies was 94.88%, the recall rate was 89.04%, and the F1 score was 91.87%.

[0147] Example 3

[0148] Combination Figure 4 This embodiment will be described in detail. The second set of parameters in this embodiment includes:

[0149] The time span of the training telemetry data is T1: 12000;

[0150] The time span for testing telemetry data is T2:8000;

[0151] The validation set partition ratio during training is 0.2.

[0152] Batch size during model training: 32;

[0153] Maximum training epochs: 30;

[0154] The time window length for telemetry historical time series slices is w: 25;

[0155] Telemetry prediction time window length p:10;

[0156] The window length for the EWMA weighted average is 50.

[0157] The proportion of early-stage non-abnormal window lengths is δ: 5%;

[0158] z:3.

[0159] The anomaly detection effect of this embodiment on satellite constellation telemetry time series data is as follows: Figure 4 As shown, the anomaly detection precision obtained on actual telemetry data with anomalies was 94.93%, the recall rate was 87.21%, and the F1 score was 90.90%.

[0160] In summary, the method proposed in this invention can guarantee both high anomaly detection accuracy and high recall, and has good application capabilities for anomaly detection in satellite constellations.

[0161] 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. A satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion, characterized in that, The steps of the satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion include: Step 1: Divide the historical satellite constellation telemetry time series data of the constellation network into training telemetry datasets and test telemetry datasets; Step 2: Preprocess the training telemetry dataset; 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: Based on the time feature extraction module, analyze the low-frequency and high-frequency components in the historical satellite constellation telemetry time series data respectively, and extract the multi-level dependency information of all telemetry channels of the satellite constellation on the historical satellite constellation telemetry time series data; Step 3.2: Based on the spatial feature extraction module, adaptively learn the adjacency matrix between all telemetry channels of the satellite constellation, determine the coupling relationship between all telemetry channels and model it, and obtain the adjacent nodes with related relationships based on the adjacency matrix, and fit and sort the current node according to the importance of the adjacent nodes with related relationships. Step 3.3: The time feature extraction module and spatial feature extraction module alternately extract information from the historical satellite constellation telemetry time series data to obtain the historical telemetry sequences of all telemetry channels in the input satellite constellation telemetry data prediction model, and output the prediction mapping function for all telemetry data in the prediction time window. ; The expressions for the low-frequency and high-frequency components are as follows: （10） （11） In formulas (10) and (11), For channel exist Low-frequency components of time, For channel exist High-frequency components at time, For channel exist Telemetry data at time Indicates average pooling; The expression for the adjacency matrix among all telemetry channels is: （12） In formula (12), E represents the learning parameters of the satellite constellation telemetry data prediction model. Transpose itself; The expression for determining the association relationship between adjacent nodes is: （13） In formula (13), For each element in the adjacency matrix A value of 1 indicates that the two nodes are related. A value of 0 indicates that the two nodes are not related. The expression for fitting and sorting adjacent nodes is: （14） （15） In formulas (14) and (15), For nodes For nodes The degree of importance, To concatenate two tensors along a common dimension, These are elements in the adjacency matrix of telemetry channel nodes. It is a normalized exponential function; The expression for the historical telemetry sequence is: （16） （17） In formulas (16) and (17), For time arrive The training dataset corresponds to the predicted telemetry time series data, time points Including time points ,time Not including time points ; For time arrive The test dataset corresponds to the predicted telemetry time series data, time Including time points ,time Not including time points ; To train time in telemetry data Time Data values, time Including time points ,time Not including time points ; To test time in telemetry data Time Data value time Including time points ,time Not including time points ; Step 4: Train the satellite constellation telemetry data prediction model based on the training telemetry dataset; Step 5: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model to obtain the satellite constellation telemetry data prediction results, and perform real-time anomaly detection on the satellite constellation telemetry data prediction; Step 5, which involves real-time anomaly detection of the satellite constellation telemetry data prediction, includes: Step 5.1: Input the test telemetry dataset into the trained satellite constellation telemetry data prediction model and calculate the absolute prediction error at each time step; Step 5.2: Calculate the smoothed prediction error of the test telemetry data using the exponentially weighted moving average method and the absolute prediction error at each time step; Step 5.3: Calculate the detection threshold for all telemetry channels based on the smoothed prediction error of the test telemetry data; Step 5.4: Perform anomaly detection on subsequent time series based on the detection thresholds of all telemetry channels. If the smoothing prediction error at a time step is greater than the detection threshold, it is considered an anomaly. The formula for calculating the detection threshold of all telemetry channels is: （25） （26） In formulas (25) and (26), To determine the length of the early anomaly-free window for anomaly detection threshold, This represents the percentage of the test data's total time span where the early, anomaly-free window is present. This is a constant between 1 and 3.

5. This parameter is a vector, and different values ​​can be selected for different channels. To test the time span of telemetry data, For the set anomaly monitoring threshold, for ~ The smoothed prediction error value, where, ~ The range includes 0 and does not include ; To calculate the mean of the data within the parentheses, To calculate the standard deviation of the data within the parentheses.

2. The satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion according to claim 1, characterized in that, In step 1, the ratio of the training telemetry dataset to the test telemetry dataset is 8:2; The expressions for the training telemetry dataset and the test telemetry dataset are as follows: （1） （2） （3） （4） （5） （6） （7） （8） In formulas (1)-(8), 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 ,time 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.

3. The satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion according to claim 1, characterized in that, Step 2 involves preprocessing the training telemetry dataset, including: The training telemetry dataset is standardized to obtain a preprocessed training telemetry dataset. The expression for standardizing the training telemetry dataset is: （9） In formula (9), 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.

4. The satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion according to claim 1, characterized in that, The satellite constellation telemetry data prediction model in step 3 includes a temporal feature extraction module and a spatial feature extraction block; The time feature extraction module is used to extract multi-level dependency information of all telemetry channels of the satellite constellation on historical satellite constellation telemetry time series data; The spatial feature extraction module is used to model all telemetry channels of the satellite constellation and to fit and sort the adjacent nodes in the telemetry channels.

5. The satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion according to claim 1, characterized in that, Step 4, which involves training the satellite constellation telemetry data prediction model, includes the following steps: Step 4.1: Input the training telemetry dataset and obtain the predicted telemetry time series data corresponding to the training telemetry dataset; Step 4.2: Calculate the root mean square error between the predicted telemetry time series data and the actual telemetry time series data corresponding to the training telemetry dataset; Step 4.3: Adjust the current training phase based on the root mean square error of the current training phase; Step 4.4: Repeat steps 4.1-4.3 until the root mean square error is less than the preset value, and obtain the trained satellite constellation telemetry data prediction model; The expression for the root mean square error loss function is: (18) In formula (18), For time arrive The training telemetry dataset corresponds to the real telemetry time series data, time Including time points ,time Not including time points ; For time arrive The training telemetry dataset corresponds to the predicted telemetry time series data, time Including time points ,time Not including time points ; The window length for splitting data during the preprocessing stage. The time span for training telemetry data.

6. The satellite constellation telemetry anomaly detection method based on adaptive spatiotemporal feature fusion according to claim 1, characterized in that, The formula for calculating absolute prediction error is: （19） （20） In formulas (19) and (20), To test the telemetry data Real data at each time step To test the telemetry data Predicted data at each time step To test the telemetry data Absolute prediction error at each time step To test the absolute prediction error of telemetry data, The window length for splitting data during the preprocessing stage. To test the time span of telemetry data; The formula for calculating the smoothing prediction error is: （21） （22） （23） （24） In formulas (21)-(24), The window length for the weighted average of the algorithm. For the smoothing parameters to satisfy This determines the weight of the most recent observation. To test the telemetry data Smoothing prediction error at each time step To test the smoothing prediction error of telemetry data.

Images