Spacecraft sensor abnormal data recovery method based on spatio-temporal information fusion

Abstract

The invention discloses a spacecraft sensor abnormal data recovery method based on spatio-temporal information fusion, and relates to a spacecraft sensor abnormal data recovery method. The method aims at solving the problems that in an existing spacecraft sensor abnormal data recovery technology, the signal-to-noise ratio is deteriorated due to environment interference, and multi-dimensional feature modeling has limitation and is insufficient in self-adaptive capacity. The method comprises the following steps: step 1, preprocessing original sensor data through a sliding mean filter; 2, constructing a time-space dual-scale feature extraction network based on Transform, mining a sensor time sequence dependency relationship from a time dimension, and analyzing multi-sensor association features from a space dimension; 3, introducing a particle filter algorithm to carry out optimal fusion on space-time dual-channel output, and quantifying the uncertainty of a data recovery result; and step 4, by comparing the actually measured data with the uncertainty confidence interval, abnormal detection and data recovery of sensor anomalies are realized. The invention belongs to the technical field of satellite sensor abnormal data recovery.

Description

Technical Field

[0001] The invention relates to a method for recovering abnormal data of a spacecraft sensor, and belongs to the technical field of abnormal data recovery of satellite sensors. Background Art

[0002] With the rapid development of aerospace and microelectronics technologies, the system architecture of modern satellites has become increasingly complex, with significantly increased functional integration. To comprehensively monitor the satellite's on-orbit operational status, a large number of high-precision payload sensors are deployed in key subsystems (including power systems, navigation systems, thermal control systems, etc.). However, under the continuous influence of the extreme space environment (such as strong radiation and drastic temperature fluctuations), coupled with inherent factors such as lithium battery capacity degradation and mechanical component fatigue and aging, payload sensors frequently experience data anomalies or even complete failure. This interfered or faulty sensor data not only seriously affects the accuracy of satellite status monitoring but also poses a severe challenge to on-orbit operation and maintenance decision-making, directly threatening the long-term reliable operation of satellites and mission stability. In this context, improving the accuracy of detecting and recovering abnormal data has become a key technical foundation for improving satellite data reliability, enhancing the accuracy of status assessments, strengthening on-orbit operation and maintenance capabilities, and ensuring the highly reliable and stable operation of satellites.

[0003] Defects and shortcomings of existing technology:

[0004] Currently, satellite sensor anomaly data recovery mainly consists of two core links: sensor data anomaly detection and data recovery. It usually builds a data recovery model based on the collected raw sensor data to achieve accurate detection and high-quality recovery of anomaly data. Existing methods mainly focus on feature analysis of data in a single time dimension or spatial dimension and have been successfully applied to traditional industrial fields such as communication systems and air handling units. However, satellite sensor data has characteristics such as high dimensionality, strong coupling, and high nonlinearity, which poses significant challenges to the construction of data recovery models. The specific challenges are as follows:

[0005] (1) SNR deterioration caused by environmental interference. Satellite sensors are exposed to extreme space environments, strong electromagnetic radiation, and unstable satellite-to-ground transmission links for a long time. These factors can introduce significant random noise and pulse interference into the raw data, seriously reducing the reliability and availability of the data.

[0006] (2) Limitations of multi-dimensional feature modeling. Most existing satellite data recovery models only consider a single dimension in time series correlation or spatial correlation, and fail to fully explore the spatiotemporal coupling relationship between high-dimensional sensor data. As a result, the constructed sensor data anomaly detection and recovery model is difficult to fully describe the change characteristics of satellite sensor data, and thus the recovery effect of anomaly data.

[0007] (3) Insufficient adaptive capabilities. Current mainstream anomaly detection methods generally use fixed threshold criteria and lack the ability to adapt to complex satellite operating conditions. This rigid discrimination mechanism severely restricts the generalization performance of data recovery methods and hinders the improvement of sensor anomaly data detection and data recovery accuracy. Summary of the Invention

[0008] In order to solve the problems of existing spacecraft sensor abnormal data recovery technology, such as deterioration of signal-to-noise ratio due to environmental interference, limitations of multi-dimensional feature modeling and insufficient adaptability, the present invention proposes a spacecraft sensor abnormal data recovery method based on spatiotemporal information fusion.

[0009] The technical solution adopted by the present invention to solve the above problems is: the steps of the present invention include:

[0010] Step 1: Preprocess the raw sensor data through a sliding mean filter to effectively suppress noise interference;

[0011] Step 2: Build a Transformer-based spatiotemporal dual-scale feature extraction network to mine sensor temporal dependencies from the temporal dimension and analyze multi-sensor correlation features from the spatial dimension, thus establishing a high-precision spatiotemporal joint data recovery model.

[0012] Step 3: Introduce the particle filter algorithm to optimally fuse the spatiotemporal dual-channel outputs and quantify the uncertainty of the data recovery results;

[0013] Step 4: By comparing the measured data with the uncertainty confidence interval, anomaly detection and data recovery of sensor anomalies are achieved.

[0014] Furthermore, step 1 specifically includes:

[0015] Obtain multi-dimensional sensor data from each satellite system

[0016]

[0017] In formula (1), M represents the dimension of sensor data, N represents the total number of points of each sensor data, Represent different types of sensor data;

[0018] According to the noise characteristics of the satellite sensor, the length of the sliding mean filter is set to L, and the data in each window is averaged; the sliding mean filter is slid on the satellite sensor time series data to obtain the denoised sensor data: D = [d1, d2, d3, ..., d N ],d i =[x i ,yi ,z i ,...,k i ],i=1,2,...,N,

[0019]

[0020] Divide the dataset D into training dataset according to 8:2: train =[d1,d2,,d N×0.8 ] with the test dataset: D test =[d N×0.8+1 ,d N×0.8+2 ,...,d N ], where the training set does not contain abnormal data and the test set contains abnormal data.

[0021] Furthermore, step 2 specifically includes:

[0022] Building a Transformer-based time-scale data recovery model

[0023]

[0024] In formula (3), the input and output dimensions of the model are the same as m;

[0025] The output of the model is:

[0026] Building a Transformer-based spatial scale data anomaly detection and recovery model

[0027] In formula (4), the input and output dimensions of the model are the same as m;

[0028] The output of the model is:

[0029] Furthermore, step 3 specifically includes:

[0030] Assume that there are γ particles, each particle represents a state, which is represented by X temporal Determine and use formula (5) and formula (6) to initialize the particles and particle weights Represents X temporal The mean of Represents X temporal variance;

[0031]

[0032] Utilize Xspatial Update particle weights and obtain observation values ​​of sensor data;

[0033]

[0034] In formula (7), r t represents noise;

[0035] when When , the observation likelihood function of satellite sensor data is as follows:

[0036]

[0037] Use formulas (7) and (8) to update and normalize;

[0038]

[0039] Repeat steps 8 to 9 to obtain the satellite sensor data with spatiotemporal information fusion: X′=[x1′,x2′,...,x′ N×0.8 ];

[0040]

[0041] Furthermore, step 4 specifically includes:

[0042] The loss function of the sensor data anomaly detection and recovery model based on spatiotemporal information fusion is designed as shown in formula (11), and the loss function is calculated to obtain the optimal spatiotemporal dual-scale fusion satellite sensor data anomaly detection and recovery model;

[0043]

[0044] Using mean square error, root mean square error and R 2 ,Evaluate the accuracy of satellite sensor data anomaly detection and recovery models;

[0045]

[0046] The test data D test Injected into the satellite sensor data anomaly detection and recovery model based on spatiotemporal information fusion, and the satellite sensor data recovery result is obtained: X′=[x′ N×0.8+1 ,x′ N×0.8+2 ,...,x′ N ]. And the particle filter is used to evaluate the uncertainty of the model results, as shown in formula (15), k can be obtained by looking up the table:

[0047]

[0048] By judging whether the observation value of satellite sensor data is within the confidence interval In the sensor data, it is determined whether an anomaly occurs. If an anomaly occurs, the output of the sensor data anomaly detection and recovery model based on spatiotemporal information fusion is used to replace the observation value to improve the reliability of the sensor data, as shown in formula (16), thereby obtaining the observation value of the satellite sensor data with high reliability and no anomaly:

[0049]

[0050] The beneficial effects of the present invention are:

[0051] 1. The present invention uses a sliding mean filtering algorithm to pre-process the raw data of spacecraft sensors, thereby reducing sensor data noise and improving the accuracy of the sensor data recovery model;

[0052] 2. This paper proposes a sensor data anomaly detection and recovery architecture based on a spatiotemporal dual-scale Transformer-particle filter. This architecture leverages the powerful spatiotemporal feature extraction capabilities of the Transformer to capture the temporal and spatial correlations of sensor data. It combines this architecture with the Bayesian inference framework of the particle filter to achieve optimal fusion of multi-source information and quantify uncertainty. Furthermore, it uses dynamic confidence intervals to replace traditional fixed thresholds, improving the adaptability and accuracy of sensor data anomaly detection.

[0053] 3. This paper develops a multi-layer spatiotemporal information fusion model that simultaneously models the spatial coupling relationship between sensors and the temporal evolution law within the sensors through a hierarchical feature extraction network, significantly improving the accuracy of sensor abnormal data recovery;

[0054] 4. Experiments were conducted using actual observation data from sensors during real satellite operation. The experimental results show that the method proposed in this paper can realize anomaly detection and high-precision recovery of satellite sensor data, thereby improving the reliability of satellite sensor data. The MSE is 0.067%, the RMSE is 2.581%, and the R 2 =98.897%. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a block diagram of the overall solution of the present invention;

[0056] Figure 2 This is a schematic diagram of the Transformer-based sensor data anomaly detection and recovery model;

[0057] Figure 3 This is a comparison chart of different methods. DETAILED DESCRIPTION

[0058] Example 1: Figures 1 to 3 As shown in FIG, a method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion includes the following steps:

[0059] Step 1: Verify using 10 sensor data including charging array current, flywheel D speed, flywheel E speed, bus output current, yaw attitude angle, solar cell array temperature, star sensor roll angle, star sensor yaw angle, battery voltage and flywheel B speed. Each sensor contains 20,000 sampling points. Therefore, the high-dimensional sensor data of each satellite system is Where, M = 10, N = 20000;

[0060] Step 2: Set the length of the sliding mean filter to L = 20, and use formula (1-2) to average the data in each window; slide the sliding mean filter on the satellite sensor time series data to obtain the denoised sensor data: D = [d1, d2, d3, ..., d N ],d i =[x i ,y i ,z i ,...,k i ],i=1,2,...,20000;

[0061] Step 3: Divide the dataset D into training set according to 8:2: D train =[d1,d2,,d 16000 ] and the test set: D test =[d 16001 ,d 16002 ,...,d 20000 ]; the training set does not contain abnormal data, and the test set contains abnormal data;

[0062] Step 4: Build a Transformer-based time-scale satellite sensor data anomaly detection and recovery model (like Figure 2 As shown in formula (1-17), the input and output dimensions of the model are the same, m = 20, and the output of the model is:

[0063]

[0064] Step 5: Build a Transformer-based spatial-scale satellite sensor data anomaly detection and recovery model (like Figure 2 As shown in formula (1-18), the input and output dimensions of the model are the same, m = 20, and the output of the model is:

[0065]

[0066] 6. Assume that there are γ = 60,000 particles, each particle represents a state, which is represented by X temporal Determine and use (1-5) and (1-6) to initialize particles and particle weights Represents X temporal The mean of Represents X temporal variance;

[0067] Step 7. Use X spatial Update the particle weights and obtain the observation values ​​of the sensor data, as shown in formula (1-7), r t represents noise; when When , the observation likelihood function of satellite sensor data is shown in formula (1-8); and formula (1-7) is used to update and normalize formula (1-8), as shown in formula (1-9); repeat (1-8) and (1-9) to obtain satellite sensor data with spatiotemporal information fusion: X′=[x′1,x′2,...,x′ 16000 ], as shown in formula (1-10); the loss function of the sensor data anomaly detection and recovery model based on spatiotemporal information fusion is designed as shown in formula (1-11), and the loss function is calculated using back propagation to obtain the optimal spatiotemporal dual-scale fusion satellite sensor data anomaly detection and recovery model;

[0068] Step 8. Compare the method proposed in the present invention with the currently commonly used sensor data anomaly detection and recovery methods (Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU) and Convolutional Neural Network (CNN)). Calculate the accuracy of the model using formulas (1-12), (1-13) and (1-14). The recovery results of different methods are compared in Table 1. Finally, the method proposed in the present invention has the highest average accuracy.

[0069] Table 1 Accuracy comparison of different methods

[0070]

[0071] Step 9: Test data D testInjected into the satellite sensor data anomaly detection and recovery model based on spatiotemporal information fusion, and the satellite sensor data recovery result is obtained: X′=[x′ 16001 ,x′ 16002 ,...,x′ 20000 ]; and use particle filtering to evaluate the uncertainty of the model results, as shown in formula (1-15), where k = 1.98 can be obtained by looking up the table;

[0072] Step 10: Finally, determine whether the observed value of the satellite sensor data is within the confidence interval. In the sensor data, it is determined whether an anomaly occurs. If an anomaly occurs, the data anomaly detection and recovery model based on spatiotemporal information fusion is used to replace the observation value to improve the reliability of the sensor data, as shown in formula (1-16), thereby obtaining the observation value of the satellite sensor data with high reliability: The method proposed in this invention is compared with the commonly used data anomaly detection and recovery methods. The results are as follows: Figure 3 As shown, it can be seen that the method proposed in the present invention has the best data recovery accuracy.

[0073] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion, characterized in that: The specific steps include: Step 1: Preprocess the raw sensor data through a sliding mean filter to effectively suppress noise interference; Step 2: Build a Transformer-based spatiotemporal dual-scale feature extraction network to mine sensor temporal dependencies from the temporal dimension and analyze multi-sensor correlation features from the spatial dimension, thus establishing a high-precision spatiotemporal joint data recovery model. Step 3: Introduce the particle filter algorithm to optimally fuse the spatiotemporal dual-channel outputs and quantify the uncertainty of the data recovery results; Step 4: By comparing the measured data with the uncertainty confidence interval, anomaly detection and data recovery of sensor anomalies are achieved.

2. The method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion according to claim 1, characterized in that: Step 1 specifically includes: Obtain multi-dimensional sensor data from each satellite system In formula (1), M represents the dimension of sensor data, N represents the total number of points of each sensor data, Represent different types of sensor data; According to the noise characteristics of the satellite sensor, the length of the sliding mean filter is set to L, and the data in each window is averaged; the sliding mean filter is slid on the satellite sensor time series data to obtain the denoised sensor data: D = [d1, d2, d3, ..., d N ],d i =[x i ,y i ,z i ,...,k i ],i=1,2,...,N, Divide the dataset D into training dataset according to 8:2: train =[d1,d2,,d N×0.8 ] with the test dataset: D test =[d N×0.8+1 ,d N×0.8+2 ,...,d N ], where the training set does not contain abnormal data and the test set contains abnormal data.

3. The method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion according to claim 1, characterized in that: Step 2 specifically includes: Building a Transformer-based time-scale data recovery model In formula (3), the input and output dimensions of the model are the same as m; The output of the model is: Building a Transformer-based spatial scale data anomaly detection and recovery model In formula (4), the input and output dimensions of the model are the same as m; The output of the model is:

4. The method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion according to claim 1, characterized in that: Step 3 specifically includes: Assume that there are γ particles, each particle represents a state, which is represented by X temporal Determine and use formula (5) and formula (6) to initialize the particles and particle weights Represents X temporal The mean of Represents X temporal variance; Utilize X spatial Update particle weights and obtain observation values ​​of sensor data; In formula (7), r t represents noise; when When , the observation likelihood function of satellite sensor data is as follows: Use formulas (7) and (8) to update and normalize; Repeat steps 8 to 9 to obtain the satellite sensor data with spatiotemporal information fusion: X′=[x1′,x2′,...,x′ N×0.8 ]; 5. The method for recovering abnormal data of spacecraft sensors based on spatiotemporal information fusion according to claim 1, characterized in that: Step 4 specifically includes: The loss function of the sensor data anomaly detection and recovery model based on spatiotemporal information fusion is designed as shown in formula (11), and the loss function is calculated to obtain the optimal spatiotemporal dual-scale fusion satellite sensor data anomaly detection and recovery model; Using mean square error, root mean square error and R 2 ,Evaluate the accuracy of satellite sensor data anomaly detection and recovery models; The test data D test Injected into the satellite sensor data anomaly detection and recovery model based on spatiotemporal information fusion, and the satellite sensor data recovery result is obtained: X′=[x′ N×0.8+1 ,x′ N×0.8+2 ,...,x′ N ]. And the particle filter is used to evaluate the uncertainty of the model results, as shown in formula (15), k can be obtained by looking up the table: By judging whether the observation value of satellite sensor data is within the confidence interval In the sensor data, it is determined whether an anomaly occurs. If an anomaly occurs, the output of the sensor data anomaly detection and recovery model based on spatiotemporal information fusion is used to replace the observation value to improve the reliability of the sensor data, as shown in formula (16), thereby obtaining the observation value of the satellite sensor data with high reliability and no anomaly:

Images