Spacecraft telemetry data anomaly detection method based on space-time correlation

Abstract

The invention provides a spacecraft telemetry data anomaly detection method based on space-time correlation, belongs to the technical field of spacecraft telemetry data anomaly detection, solves the problems of low accuracy and high false alarm rate of a spacecraft telemetry data anomaly detection result, and comprises the following steps: 1, obtaining spacecraft telemetry data; 2, preprocessing the telemetry data of the spacecraft; 3, constructing a feature extraction model, training the feature extraction model, and carrying out feature extraction on the preprocessed spacecraft telemetry data based on the trained feature extraction model; 4, carrying out the correlation fusion of the feature extraction results, and obtaining a spacecraft telemetry data prediction result of the t time step; and step 5, obtaining an abnormal threshold based on the training data of the feature extraction model, calculating an abnormal score of the spacecraft telemetry data at the time step t, if the abnormal score of the time step t deviates from the abnormal threshold, determining that the data is abnormal data, and completing the anomaly detection of the spacecraft telemetry data.

Description

Technical Field

[0001] The invention relates to a spacecraft telemetry data anomaly detection method based on time-space correlation, belonging to the technical field of spacecraft telemetry data anomaly detection. Background Art

[0002] Spacecraft telemetry data refers to the process during the spacecraft's on-orbit operation when the sensors in its telemetry system convert the measured data into electrical signals. Each electrical signal is transmitted to the ground telemetry equipment. The ground equipment uses signal demodulation technology to restore the original parameter information of each channel, which is the spacecraft telemetry data. Due to the uncertainty of the spacecraft's mission and the various types of failures that may occur in the spacecraft itself, the spacecraft may generate unexpected abnormal data during operation, which may cause the spacecraft's operation mode to be abnormal or even cause serious losses. For example, the Worldview-4 satellite developed by Digital Globe in the United States eventually caused a loss of nearly $144 million due to abnormal attitude and orbit control telemetry data; the Hitomi satellite developed by JAXA in conjunction with NASA and other research institutions had a malfunction in its telemetry data, resulting in the breakage and damage of components such as solar panels. Therefore, in order to ensure the safe and stable operation of the spacecraft, it is necessary to detect possible anomalies in its telemetry data in real time to provide information support for abnormal source identification and reliable decision-making, thereby further improving its operational safety.

[0003] Most of the existing anomaly detection methods only consider the correlation between sensor data or the correlation change between different sensor data. When the spacecraft operating conditions change dynamically, the accuracy of anomaly detection results may be low and the false alarm rate may be high due to insufficient feature extraction of sensor data. The specific challenges include: (1) From the perspective of model selection, most of the existing methods are designed based on models such as relevance vector machine (RVM), long short-term memory network (LStM), gated recurrent unit (GRU). However, due to the limited number of network layers, such methods may cause insufficient feature extraction. Although networks such as LStM introduce selective memory and forgetting mechanisms, they may still miss historical data information when extracting features related to temporal evolution or spatial changes of complex data before and after; (2) From the perspective of method establishment, most methods only consider the temporal evolution characteristics of spacecraft telemetry data or the spatial correlation characteristics between different sensors for anomaly detection. When the mission mode and operation status of a spacecraft present a dynamic change trend, the correlation characteristics of the same sensor data before and after and the correlation characteristics between different sensor data may not be significant. Modeling only through temporal evolution in the time dimension or data correlation modeling in the space dimension is prone to the problem of low accuracy and high false alarm rate in the anomaly detection results due to the difficulty in fully expressing the change trend of the data. In view of the above problems, the present invention proposes a method for detecting anomaly in spacecraft telemetry data based on time-space correlation. Summary of the invention

[0004] In order to solve the above-mentioned problems in the anomaly detection of spacecraft telemetry data, the present invention further proposes a method for anomaly detection of spacecraft telemetry data based on time-space correlation.

[0005] The technical solution adopted by the present invention to solve the above problems is: the present invention proposes a method for detecting anomalies in spacecraft telemetry data based on time-space correlation, comprising:

[0006] Step 1: Obtain spacecraft telemetry data;

[0007] Step 2: Preprocess the spacecraft telemetry data;

[0008] Step 3: Construct a feature extraction model and train the feature extraction model, and extract features from the preprocessed spacecraft telemetry data based on the trained feature extraction model;

[0009] Step 4: Correlate and fuse the feature extraction results to obtain the prediction results of the spacecraft telemetry data at time step t;

[0010] Step 5: Obtain the anomaly threshold based on the training data of the feature extraction model, and calculate the anomaly score of the spacecraft telemetry data at time step t. If the anomaly score at time step t deviates from the anomaly threshold, it is anomaly data, and the anomaly detection of spacecraft telemetry data is completed.

[0011] Optionally, the spacecraft telemetry data in step 1 is an ordered set of elements consisting of record values ​​and record times;

[0012] The expression of the data sequence of spacecraft i is:

[0013] S i =[s i1 ,s i2 ,...,s it ](1);

[0014] In formula (1), s it is the telemetry value of spacecraft i at time t.

[0015] Optionally, the step of preprocessing the spacecraft telemetry data in step 2 includes:

[0016] Step 2.1: For the spacecraft telemetry data sequence S of the time dimension m to be measured m and the spacecraft telemetry data sequence S of the relevant dimension n Perform normalization processing;

[0017] Step 2.2: Divide the normalized spacecraft telemetry data into time windows of length t, and put the normalized spacecraft telemetry data into the corresponding time windows;

[0018] The calculation formula for normalization is:

[0019]

[0020] In formula (2), S i is the original data sequence of the spacecraft i dimension, S min and S max S i The minimum and maximum values ​​of .

[0021] Optionally, the feature extraction model in step 3 includes a temporal evolution feature extraction module and a spatial correlation feature extraction module;

[0022] The temporal evolution feature extraction module uses a TCN network, and the residual block structure is used for skip connection between two adjacent layers of the TCN network to extract the time dimension features of the spacecraft telemetry data sequence to be detected with dimension m;

[0023] The temporal evolution feature extraction module and the spatial correlation feature extraction module use the TCN network to extract the spacecraft telemetry data sequence S of the dimension to be detected. n Perform feature extraction in the spatial dimension.

[0024] Optionally, the step of extracting features of the time dimension of the spacecraft telemetry data sequence to be detected in dimension m includes:

[0025] Step 3.1.1: Each layer of the TCN network is based on the input spacecraft telemetry data sequence S of the dimension m to be detected m Perform adaptive adjustment of the receptive field to obtain the spacecraft telemetry data sequence after dilation convolution of each channel in each layer of the TCN network;

[0026] Step 3.1.2: Introduce 1×1 convolution to make the number of feature map channels added on both sides of the skip connection layer consistent;

[0027] Step 3.1.3: Use the linear rectification function to activate the feature output of the residual fast, and obtain the spacecraft telemetry data sequence S with a dimension m to be predicted for each channel in each layer of the TCN network m The feature output of time dimension is completed;

[0028] The spacecraft telemetry data sequence S after each channel dilation convolution m The expression is:

[0029]

[0030] In formula (3), f is the convolution filter, k is the kernel size, d is the dilation factor, which indicates the number of intervals between each element of the convolution kernel, and td·i is the number of times only S is used. i The past time state is convolved. During the convolution process, the output at time t is only related to the convolution of time t and earlier elements in the previous layer;

[0031] The expression of the linear rectification function is:

[0032] ReLU(s)=max(0,s)(4);

[0033] Spacecraft telemetry data S after time dimension feature extraction m The expression of the characteristic output is:

[0034]

[0035] In formula (5), is the output of the spacecraft in the dimension to be measured predicted by each layer of the feature extraction module, TCN (S m ) is for S m The feature output obtained by forward reasoning using the TCN network, F(Sm ) is for S m Perform hole convolution and activation function operations, C 1x1 (S m ) is for S m Perform a 1×1 convolution.

[0036] Optionally, the spacecraft telemetry data sequence S of the dimension to be detected n The feature extraction of spatial dimension specifically includes:

[0037] The spacecraft telemetry data sequence S of the relevant dimension n As the independent variable of the regression analysis, the spacecraft telemetry data sequence S of the time dimension m to be measured is m As the dependent variable of the regression analysis, the TCN network is used for regression analysis to obtain the spacecraft telemetry data sequence S of the relevant dimension n =[s n1 ,s n2 ,...,s nt ]The output of each channel in each layer of TCN network after spatial correlation feature extraction;

[0038] Spacecraft telemetry data sequence S after time dimension feature extraction n The expression of the characteristic output is:

[0039]

[0040] In formula (6), is the spacecraft output of the measured dimension predicted by the spatial variation correlation feature extraction module, TCN (S n ) is for S n Using the feature output obtained by TCN network forward reasoning, F(S n ) is for S n Perform hole convolution and activation function operations, C 1x1 (S n ) is for S n Perform a 1×1 convolution.

[0041] Optionally, the steps of obtaining the predicted results of the spacecraft telemetry data at time step t in step 4 include:

[0042] Step 4.1: Calculate the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module according to the optimal loss function value obtained during the historical training of the temporal evolution feature extraction module and the spatial correlation feature extraction module;

[0043] Step 4.2: Based on the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module, the prediction outputs of the temporal evolution feature extraction module and the spatial correlation feature extraction module are associated and fused to obtain the prediction output at time step t.

[0044] The expressions of the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module are:

[0045]

[0046] In formulas (7) and (8), loss time 、loss space They are the optimal loss function value of the temporal evolution feature extraction module and the optimal loss function value of the spatial correlation feature extraction module respectively;

[0047] The predicted output at time step t is The expression is:

[0048]

[0049] Optionally, the step of detecting anomalies in spacecraft telemetry data in step 5 includes:

[0050] Step 5.1: Calculate the mean μ and standard deviation σ of the normal data and predicted data during feature extraction module training;

[0051] Step 5.2: Calculate the anomaly threshold λ based on the mean μ and standard deviation. The upper bound of the anomaly threshold λ is λ up , the lower limit of the threshold is λ low ;

[0052] Step 5.3: Calculate the anomaly score of the spacecraft telemetry data of the measured dimension m at time step t

[0053] Step 5.4: Set the anomaly detection label set ξ=[ξ1,ξ2,...,ξ t ], if the anomaly score at time step t If it deviates from the set abnormal threshold λ, the spacecraft telemetry data of the corresponding time step is judged as abnormal data, and the corresponding abnormal detection label is marked as 1. If it is within the abnormal threshold λ, it is normal data, and the corresponding abnormal detection label is marked as 0, completing the abnormal detection of spacecraft telemetry data;

[0054] Abnormal threshold upper bound λ up and the threshold lower bound is λ low The calculation formulas are:

[0055] λ up =μ+3σ (10);

[0056] λ low =μ-3σ (11);

[0057] The anomaly score of the spacecraft telemetry data of the measured dimension m at time step t The calculation formula is:

[0058]

[0059] In formula (12), s mt is the predicted value s t The corresponding raw data observation, s mt The dimension is m and the step size is t.

[0060] The beneficial effects of the present invention are:

[0061] 1. The present invention adopts TCN network to design the spacecraft telemetry data feature extraction module. Due to its flexible receptive field, the network can better extract global information, so it can flexibly and efficiently extract the time series evolution characteristics of complex spacecraft telemetry data or the data correlation characteristics between sensors;

[0062] 2. The present invention realizes the parallel extraction of the time and space dimensional features of spacecraft data based on the time-space correlation method, which more comprehensively expresses the changing characteristics of spacecraft telemetry data, thereby improving the accuracy and reducing the false alarm rate when performing anomaly detection of spacecraft telemetry data.

[0063] 3. The present invention takes satellites as one of the target application objects, and verifies and evaluates the proposed method by injecting data of abnormal modes such as jamming and constant deviation that may exist in the attitude and orbit control system. The results show that the present invention achieves high accuracy (99.6%) and low false alarm rate (1.5%) abnormal detection of different types of faults of different types of spacecraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A flowchart of a method for detecting anomaly in spacecraft telemetry data based on time-space correlation provided by the present invention;

[0065] Figure 2 A roll angular rate data distribution diagram of the attitude and orbit control system provided by the present invention;

[0066] Figure 3 This is a distribution diagram of abnormal scores of roll angular rate abnormal data provided by the present invention. DETAILED DESCRIPTION

[0067] Combination Figure 1-3 This embodiment is described as follows. Figure 1As shown, the steps of a method for detecting anomaly of spacecraft telemetry data based on time-space correlation described in this embodiment include:

[0068] S1: The method is verified using the telemetry data of the satellite attitude and orbit control system as a representative, and fault injection is performed on the real roll angular rate dimension data. A total of 5000 sets of normal and abnormal data containing faults are as follows Figure 2 shown.

[0069] S2: The telemetry data sequence of the satellite attitude and orbit control system for the dimension to be measured, that is, the roll angular rate sequence S rr And the spacecraft telemetry data sequence of the remaining dimensions of the IMU, that is, the x-axis acceleration sequence S ax , y-axis acceleration sequence S ay , z-axis acceleration sequence S az , pitch angle rate sequence S pr , yaw rate sequence S yr All data are normalized, and the normalized data are divided into time windows of length t and enter the window to be detected so that the feature extraction model can process the data.

[0070] The calculation formula for normalization is:

[0071]

[0072] In formula (1), S i is the original data sequence of the spacecraft i dimension, S min and S max S i The minimum and maximum values ​​of .

[0073] S3: Use the normal mode satellite attitude and orbit control system telemetry data and the abnormal data containing faults to train the feature extraction module, and obtain the optimal loss function value loss obtained in the historical training process time 、loss space ;

[0074] S4: A 5-layer TCN network is used to construct a time series evolution feature extraction module to extract the rolling angle rate sequence S rr Perform feature extraction in the time dimension, and after the dilated convolution layer, obtain the spacecraft telemetry data sequence S after dilated convolution of each channel m ;

[0075] The spacecraft telemetry data sequence S after each channel dilation convolution m The expression is:

[0076]

[0077] In formula (2), f is the convolution filter, k is the kernel size, d is the dilation factor, which indicates the number of intervals between each element of the convolution kernel, and td·i is the number of times only S is used. i The past time state is convolved. During the convolution process, the output at time t is only related to the convolution of time t and earlier elements in the previous layer.

[0078] S401: For a 5-layer TCN network, a residual block structure is used to perform jump connections between two adjacent layers of the TCN network. At the same time, a 1×1 convolution is introduced to make the number of channels of the added feature maps consistent.

[0079] S402: Use the linear rectification function (ReLU function) to activate the feature output of the residual block, and finally obtain the time series evolution feature extraction module for each channel of each layer for the roll angle rate sequence S rr The characteristic output of

[0080] The expression of the linear rectification function is:

[0081] ReLU(s)=max(0,s)(3);

[0082] Each channel of each layer of the time series evolution feature extraction module is used for the roll angular rate sequence S rr The characteristic output is:

[0083]

[0084] S5: Using TCN network, construct the rolling angle rate sequence S of the spatial correlation feature extraction module rr Extract the features of the spatial dimension. ax , y-axis acceleration sequence S ay , z-axis acceleration sequence S az , pitch angle rate sequence S pr , yaw rate sequence S yr As the regression independent variable, the roll angular rate S rr As the dependent variable of regression analysis, the regression analysis is performed. For the input satellite attitude and orbit control system telemetry data sequence S rr After the spatial correlation feature extraction, the output of each channel of each layer of the spatial correlation feature extraction module is:

[0085]

[0086] The present invention adopts the TCN network to design a spacecraft telemetry data feature extraction module. Due to its flexible receptive field, the network can better extract global information, so it can flexibly and efficiently extract the temporal evolution characteristics of complex spacecraft telemetry data or the data correlation characteristics between sensors. In addition, the present invention realizes the parallel extraction of time and space dimensional features of spacecraft telemetry data based on the time-space association method, which more comprehensively expresses the change characteristics of spacecraft telemetry data, thereby improving the accuracy and reducing the false alarm rate when performing spacecraft telemetry data anomaly detection.

[0087] S6: The optimal loss function value loss obtained based on the historical training process of the time and space modules time 、loss space , respectively calculate the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module. After calculation, the weights of the two are approximately 0.91 and 0.09 respectively;

[0088] The expressions of the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module are:

[0089]

[0090] In formulas (6) and (7), loss time 、loss space They are the optimal loss function values ​​of the temporal evolution feature extraction module and the optimal loss function values ​​of the spatial correlation feature extraction module, respectively.

[0091] S7: Correlate and fuse the prediction outputs of the two feature extraction modules to obtain the roll angular rate sequence S rr The predicted output at time step t

[0092] The predicted output at time step t is The expression is:

[0093]

[0094] S8: Calculate the mean μ and standard deviation σ of normal data and predicted data during feature extraction module training, calculate the abnormal threshold based on the mean μ and standard deviation, and perform the upper limit λ of the abnormal judgment threshold up With the lower threshold λ low After calculation, the roll angular rate sequence S rr The upper and lower thresholds for abnormality determination are 0.10 and -0.12 respectively;

[0095] Abnormal threshold upper bound λ up and the threshold lower bound is λ low The calculation formulas are:

[0096] λ up =μ+3σ (9);

[0097] λ low =μ-3σ (10).

[0098] S9: Calculate the anomaly score of the spacecraft telemetry data of the measured dimension m at time step t After calculation, the abnormal score distribution of 5000 sets of abnormal roll rate data with faults is as follows: Figure 3 As shown;

[0099] The anomaly score of the spacecraft telemetry data of the measured dimension m at time step t The calculation formula is:

[0100]

[0101] In formula (11), s mt is the predicted value s t The corresponding raw data observation, s mt The dimension is m and the step size is t.

[0102] S10: Set the predicted abnormal label set ξ=[ξ1,ξ2,...,ξ t ], if the anomaly score at time step t If the spacecraft telemetry data deviates from the set abnormal threshold λ, the time step is judged as abnormal data. Correspondingly, the abnormal label ξ corresponding to time step t is t If the label set is positive, it is marked as 1, otherwise it is marked as 0. The predicted data anomaly label set is compared with the actual data anomaly label set, and the accuracy (ACC), false alarm rate (FAR) and F1 score (F1) are calculated. The calculation results are shown in Table 1.

[0103] Table 1

[0104]

[0105] In summary, the present invention takes spacecraft as one of the target application objects, and verifies and evaluates the proposed method by injecting data of abnormal modes such as jamming and constant deviation that may exist in the attitude and orbit control system. The results show that the present invention achieves high accuracy (99.6%) and low false alarm rate (1.5%) abnormal detection of different types of faults in different types of spacecraft.

[0106] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting anomaly in spacecraft telemetry data based on spatiotemporal correlation, characterized in that: The steps of the spacecraft telemetry data anomaly detection method based on time-space correlation include: Step 1: Obtain spacecraft telemetry data; Step 2: preprocessing the spacecraft telemetry data; Step 3: Construct a feature extraction model and train the feature extraction model, and extract features from the preprocessed spacecraft telemetry data based on the trained feature extraction model; Step 4: Correlate and fuse the feature extraction results to obtain the prediction results of the spacecraft telemetry data at time step t; Step 5: Obtain the anomaly threshold based on the training data of the feature extraction model, and calculate the anomaly score of the spacecraft telemetry data at time step t. If the anomaly score at time step t deviates from the anomaly threshold, it is anomaly data, and the anomaly detection of spacecraft telemetry data is completed.

2. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 1, characterized in that: In step 1, the spacecraft telemetry data is an ordered set of elements consisting of recording values ​​and recording times; The expression of the data sequence of spacecraft i is: S i =[s i1 ,s i2 ,...,s it ](1); In formula (1), s it is the telemetry value of spacecraft i at time t.

3. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 1, characterized in that: The steps for preprocessing spacecraft telemetry data in step 2 include: Step 2.1: For the spacecraft telemetry data sequence S of the time dimension m to be measured m and the spacecraft telemetry data sequence S of the relevant dimension n Perform normalization processing; Step 2.2: Divide the normalized spacecraft telemetry data into time windows of length t, and put the normalized spacecraft telemetry data into the corresponding time windows; The calculation formula for normalization is: In formula (2), S i is the original data sequence of the spacecraft i dimension, S min and S max S i The minimum and maximum values ​​of .

4. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 1, characterized in that: The feature extraction model in step 3 includes a temporal evolution feature extraction module and a spatial correlation feature extraction module; The temporal evolution feature extraction module adopts a TCN network, and a residual block structure is used to perform jump connection between two adjacent layers of the TCN network, so as to extract the time dimension features of the spacecraft telemetry data sequence to be detected with a dimension m; The temporal evolution feature extraction module and the spatial correlation feature extraction module use a TCN network to extract the spacecraft telemetry data sequence S of the dimension to be detected. n Perform feature extraction in the spatial dimension.

5. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 4, characterized in that: The steps of extracting the time dimension features of the spacecraft telemetry data sequence to be detected in dimension m include: Step 3.1.1: Each layer of the TCN network is based on the input spacecraft telemetry data sequence S of the dimension m to be detected m Perform adaptive adjustment of the receptive field to obtain the spacecraft telemetry data sequence after dilation convolution of each channel in each layer of the TCN network; Step 3.1.2: Introduce 1×1 convolution to make the number of feature map channels added on both sides of the skip connection layer consistent; Step 3.1.3: Use the linear rectification function to activate the feature output of the residual fast, and obtain the spacecraft telemetry data sequence S with a dimension m to be predicted for each channel in each layer of the TCN network m The feature output of time dimension is completed; The spacecraft telemetry data sequence S after each channel dilation convolution m The expression is: In formula (3), f is the convolution filter, k is the kernel size, d is the dilation factor, which indicates the number of intervals between each element of the convolution kernel, and td·i is the number of times only S is used. i The past time state is convolved. During the convolution process, the output at time t is only related to the convolution of time t and earlier elements in the previous layer; The expression of the linear rectification function is: ReLU(s)=max(0,s)(4); Spacecraft telemetry data S after time dimension feature extraction m The expression of the characteristic output is: In formula (5), is the output of the spacecraft in the dimension to be measured predicted by each layer of the feature extraction module, TCN (S m ) is for S m The feature output obtained by forward reasoning using the TCN network, F(S m ) is for S m Perform hole convolution and activation function operations, C 1x1 (S m ) is for S m Perform a 1×1 convolution.

6. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 4, characterized in that: The spacecraft telemetry data sequence S of the dimension to be detected n The feature extraction of spatial dimension specifically includes: The spacecraft telemetry data sequence S of the relevant dimension n As the independent variable of the regression analysis, the spacecraft telemetry data sequence S of the time dimension m to be measured is m As the dependent variable of the regression analysis, the TCN network is used for regression analysis to obtain the spacecraft telemetry data sequence S of the relevant dimension n =[s n1 ,s n2 ,...,s nt ]The output of each channel in each layer of TCN network after spatial correlation feature extraction; Spacecraft telemetry data sequence S after time dimension feature extraction n The expression of the characteristic output is: In formula (6), is the spacecraft output of the measured dimension predicted by the spatial variation correlation feature extraction module, TCN (S n ) is for S n Using the feature output obtained by TCN network forward reasoning, F(S n ) is for S n Perform hole convolution and activation function operations, C 1x1 (S n ) is for S n Perform a 1×1 convolution.

7. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 1, characterized in that: The steps for obtaining the predicted results of the spacecraft telemetry data at time step t in step 4 include: Step 4.1: Calculate the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module according to the optimal loss function value obtained during the historical training of the temporal evolution feature extraction module and the spatial correlation feature extraction module; Step 4.2: Based on the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module, the prediction outputs of the temporal evolution feature extraction module and the spatial correlation feature extraction module are associated and fused to obtain the prediction output at time step t. The expressions of the weight w1 of the temporal evolution feature extraction module and the weight w2 of the spatial correlation feature extraction module are: In formulas (7) and (8), loss time 、loss space They are the optimal loss function value of the temporal evolution feature extraction module and the optimal loss function value of the spatial correlation feature extraction module respectively; The predicted output at time step t is The expression is:

8. The method for detecting anomaly of spacecraft telemetry data based on time-space correlation according to claim 1, characterized in that: The steps of spacecraft telemetry data anomaly detection in step 5 include: Step 5.1: Calculate the mean μ and standard deviation σ of the normal data and predicted data during feature extraction module training; Step 5.2: Calculate the anomaly threshold λ based on the mean μ and standard deviation. The upper bound of the anomaly threshold λ is λ up , the lower limit of the threshold is λ low ; Step 5.3: Calculate the anomaly score of the spacecraft telemetry data of the measured dimension m at time step t Step 5.4: Set the anomaly detection label set ξ=[ξ1,ξ2,...,ξ t ], if the anomaly score at time step t If it deviates from the set abnormal threshold λ, the spacecraft telemetry data of the corresponding time step is judged as abnormal data, and the corresponding abnormal detection label is marked as 1. If it is within the abnormal threshold λ, it is normal data, and the corresponding abnormal detection label is marked as 0, completing the abnormal detection of spacecraft telemetry data; Abnormal threshold upper bound λ up and the threshold lower bound is λ low The calculation formulas are: l up =μ+3σ (10); l low =μ-3σ (11); The anomaly score of the spacecraft telemetry data of the measured dimension m at time step t The calculation formula is: In formula (12), s mt is the predicted value s t The corresponding raw data observation, s mt The dimension is m and the step size is t.

Images