Abnormality detection method and system for spacecraft attitude control system

Abstract

The present application relates to a method and system for detecting anomalies in a spacecraft attitude control system. The method includes: constructing a conditional variational autoencoder-gated recurrent unit network model; for the conditional variational autoencoder-gated recurrent unit network model, constructing an online optimizer for a fault feature extraction model, obtaining telemetry parameters and inputting them into the conditional variational autoencoder-gated recurrent unit network model after updating weight parameters, obtaining the fault features output by the conditional variational autoencoder-gated recurrent unit network model and calculating anomaly scores based on the fault features, and outputting a comparison result between the anomaly scores and a preset threshold. The present application adopts the above method, obtains a feature extraction model by constructing a variational autoencoder-gated recurrent unit network; designs a model optimizer, and updates the model weight parameters online; and uses a streaming data peak algorithm to calculate and update adaptive thresholds in real time to achieve real-time dynamic detection of anomalies.

Description

Technical Field

[0001] The present application relates to the technical field of spacecraft anomaly detection, and in particular to a method and system for detecting anomalies in a spacecraft attitude control system. Background Art

[0002] With the rapid development of aerospace technology, the number and complexity of satellites in orbit around the world are increasing day by day. However, satellites are in the harsh environment of space for a long time and are vulnerable to solar flares, magnetic field disturbances, electrostatic discharges, etc., which may lead to malfunctions or even failures. This poses a great challenge to the long-term operation of satellites in orbit.

[0003] Fault diagnosis technology can effectively help staff judge the operating status of on-orbit satellites, reduce the operating costs of satellites, and improve the ability of satellites to complete on-orbit tasks; artificial intelligence-based methods are the most widely used and most researched direction in the field of fault diagnosis. This method usually uses a large amount of data to train the network model, learn the nonlinear mapping in the fitting data, and obtain the fault diagnosis model.

[0004] In the prior art, since on-orbit fault data is scarce and the distribution characteristics of on-orbit data parameters change dynamically with the space environment, the accuracy of fault detection is low and false alarms are prone to occur, resulting in a high false alarm rate. Summary of the invention

[0005] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium and computer program product for detecting anomalies in a spacecraft attitude control system, which can solve the problem of scarce on-orbit fault data and dynamic changes in parameter distribution characteristics with the space environment, in response to the above-mentioned technical problems.

[0006] In a first aspect, the present application provides a method for detecting anomalies in a spacecraft attitude control system. The method comprises:

[0007] Construct a conditional variational autoencoder-gated recurrent unit network model;

[0008] For the conditional variational autoencoder-gated recurrent unit network model, constructing a fault feature extraction model online optimizer, wherein the fault feature extraction model online optimizer is used to update weight parameters of the conditional variational autoencoder-gated recurrent unit network model;

[0009] The telemetry parameters are obtained and input into the conditional variational autoencoder-gated recurrent unit network model after the weight parameters are updated, the fault features output by the conditional variational autoencoder-gated recurrent unit network model are obtained and the anomaly score is calculated based on the fault features, and the comparison result between the anomaly score and the preset threshold is output.

[0010] In one embodiment, constructing a conditional variational autoencoder-gated recurrent unit network model includes:

[0011] Selecting a corresponding standard threshold range based on the physical characteristics of the telemetry parameters, and eliminating outliers of the telemetry parameters based on the standard threshold range;

[0012] The telemetry parameters after outliers are removed are used to train a reconstruction model and a prediction model respectively, wherein the reconstruction model is a conditional variational autoencoder and the prediction model is a gated recurrent unit network;

[0013] Reconstruct and predict the on-orbit data based on the reconstruction model and the prediction model and obtain the corresponding prediction value;

[0014] The fault characteristics are calculated based on the predicted values ​​and the corresponding measured values.

[0015] In one embodiment, training the reconstruction model with the selected telemetry parameters separately includes:

[0016] Get telemetry parameters of a sliding window of preset length;

[0017] The label corresponding to the majority of the motor mode flag bits in the window is set as the window label;

[0018] The actuator model, dynamics model and kinematics model based on the conditional variational autoencoder-gated recurrent unit network model are constructed respectively;

[0019] The actuator model, dynamic model and kinematic model are trained based on telemetry parameters and window labels respectively.

[0020] In one embodiment, using the telemetry parameters after outliers are removed to train the prediction model includes:

[0021] constructing a prediction model and inputting telemetry parameters of a time sliding window of a preset length into the prediction model;

[0022] Predicting the flywheel speed at a future time based on the historical flywheel motor current, historical control torque and historical flywheel speed in the historical telemetry parameters;

[0023] Predict the flywheel angular velocity at the future time based on the historical flywheel speed and historical attitude angular velocity in the historical telemetry parameters;

[0024] The flywheel attitude angle at the future moment is predicted based on the historical angular velocity and historical attitude angle in the historical telemetry parameters.

[0025] In one embodiment, a verification operation is performed based on the reconstructed values ​​and true values ​​corresponding to the reconstructed model and the predicted model:

[0026] Calculate the mean square error based on the reconstructed value and the true value;

[0027] Calculate the goodness of fit based on the reconstructed and true values.

[0028] In one embodiment, for a conditional variational autoencoder-gated recurrent unit network model, building an online optimizer for a fault feature extraction model includes:

[0029] Get the loss function corresponding to the conditional variational autoencoder-gated recurrent unit network model;

[0030] Calculate the cumulative gradient of the loss function of the preset rounds and the corresponding regularized hyperparameters to obtain the weight update formula;

[0031] The updated values ​​of the parameters of each dimension are calculated based on the optimizer hyperparameters and parameter cumulative gradients.

[0032] In a second aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0033] Construct a conditional variational autoencoder-gated recurrent unit network model;

[0034] For the conditional variational autoencoder-gated recurrent unit network model, constructing a fault feature extraction model online optimizer, wherein the fault feature extraction model online optimizer is used to update weight parameters of the conditional variational autoencoder-gated recurrent unit network model;

[0035] The telemetry parameters are obtained and input into the conditional variational autoencoder-gated recurrent unit network model after the weight parameters are updated, the fault features output by the conditional variational autoencoder-gated recurrent unit network model are obtained and the anomaly score is calculated based on the fault features, and the comparison result between the anomaly score and the preset threshold is output.

[0036] In a third aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0037] Construct a conditional variational autoencoder-gated recurrent unit network model;

[0038] For the conditional variational autoencoder-gated recurrent unit network model, constructing a fault feature extraction model online optimizer, wherein the fault feature extraction model online optimizer is used to update weight parameters of the conditional variational autoencoder-gated recurrent unit network model;

[0039] The telemetry parameters are obtained and input into the conditional variational autoencoder-gated recurrent unit network model after the weight parameters are updated, the fault features output by the conditional variational autoencoder-gated recurrent unit network model are obtained and the anomaly score is calculated based on the fault features, and the comparison result between the anomaly score and the preset threshold is output.

[0040] In a fourth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0041] Construct a conditional variational autoencoder-gated recurrent unit network model;

[0042] For the conditional variational autoencoder-gated recurrent unit network model, constructing a fault feature extraction model online optimizer, wherein the fault feature extraction model online optimizer is used to update weight parameters of the conditional variational autoencoder-gated recurrent unit network model;

[0043] The telemetry parameters are obtained and input into the conditional variational autoencoder-gated recurrent unit network model after the weight parameters are updated, the fault features output by the conditional variational autoencoder-gated recurrent unit network model are obtained and the anomaly score is calculated based on the fault features, and the comparison result between the anomaly score and the preset threshold is output.

[0044] The above-mentioned spacecraft attitude control system anomaly detection method, computer equipment, storage medium and computer program product construct a conditional variational autoencoder-gated recurrent unit network model; for the conditional variational autoencoder-gated recurrent unit network model, construct a fault feature extraction model online optimizer, and the fault feature extraction model online optimizer is used to update the weight parameters of the conditional variational autoencoder-gated recurrent unit network model; obtain telemetry parameters and input them into the conditional variational autoencoder-gated recurrent unit network model after updating the weight parameters, obtain the fault features output by the conditional variational autoencoder-gated recurrent unit network model and calculate the anomaly score based on the fault features, and output the comparison result between the anomaly score and the preset threshold. This application adopts the above-mentioned method, obtains the feature extraction model by constructing a CVAE-GRU network; and designs a model optimizer to update the model weight parameters online; then uses the streaming data peak algorithm to calculate and update the adaptive threshold in real time to achieve real-time dynamic detection of anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A diagram of a spacecraft attitude control system anomaly detection scenario in one embodiment;

[0046] Figure 2 A schematic diagram of a flow chart of a method for detecting anomalies in a spacecraft attitude control system in one embodiment;

[0047] Figure 3 A network structure diagram of CAVE in one embodiment;

[0048] Figure 4 A GRU network structure diagram in one embodiment;

[0049] Figure 5 This is a diagram of the internal structure of a GRU unit in one embodiment;

[0050] Figure 6 A schematic diagram of CAVE model input and output in one embodiment;

[0051] Figure 7 A schematic diagram of input and output of a GRU model in one embodiment;

[0052] Figure 8 is a reconstruction result of the flywheel current by the CAVE model in one embodiment;

[0053] Fig. 9 A reconstructed structure of the control torque by a CAVE model in one embodiment;

[0054] Fig.10 is a result of reconstructing the flywheel speed using a CAVE model in one embodiment;

[0055] Fig.11 is a reconstruction result of CVAE for attitude roll angular velocity, attitude pitch angular velocity, and attitude yaw angular velocity in one embodiment;

[0056] Fig.12 is a prediction result of the flywheel speed by GRU in one embodiment;

[0057] Fig.13 is the prediction result of the GRU for the attitude roll angular velocity, attitude pitch angular velocity, and attitude yaw angular velocity in one embodiment;

[0058] Fig.14 is the prediction result of the GRU for the attitude roll angle, attitude pitch angle, and attitude yaw angle in one embodiment;

[0059] Fig.15 A feature extraction scheme based on CVAE-GRU in one embodiment;

[0060] Fig.16 This is a comparison diagram of the effects before and after the preset value is removed in an embodiment;

[0061] Fig.17 This is a comparison diagram of the effects before and after resampling in one embodiment;

[0062] Fig.18 A comparison diagram of the effects before and after missing value completion in an embodiment;

[0063] Fig.19 1. Anomaly score ROC curve diagram of three feature extraction models in one embodiment;

[0064] Fig. 20 A comparison diagram of the effects of anomaly detection of an FTRL feature extraction optimizer in one embodiment;

[0065] Fig.21 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0067] The method for detecting anomalies in a spacecraft attitude control system provided in the embodiment of the present application can be applied to Figure 1 In the abnormal detection scenario of the spacecraft attitude control system shown in the figure, the spacecraft sends the real-time status data to the data support software, the data support software sends the real-time status data of the whole satellite to the onboard software, and sends the function switch and algorithm parameters and other data to the onboard software in the form of software instructions. The onboard software sends the algorithm results and software running status to the data support software in the form of software instructions based on the real-time status data of the whole satellite. The ground workstation obtains the algorithm results and software running status obtained by the data support software in the form of software telemetry, and the ground workstation sends the function switch and algorithm parameters to the data support software in the form of software instructions.

[0068] In one embodiment, Figure 2 As shown, in this embodiment, the method includes the following steps:

[0069] Step 202: construct a conditional variational autoencoder-gated recurrent unit network model.

[0070] Among them, the reconstruction and prediction methods are combined to design a fault feature extraction method based on the conditional variational autoencoder-gated recurrent unit network (CVAE-GRU) model. This method first designs the CVAE-GRU network structure and uses a multi-layer neural network model to amplify the fault features layer by layer.

[0071] Step 204: construct an online optimizer for the fault feature extraction model for the conditional variational autoencoder-gated recurrent unit network model using the FTRL (Follow The Regularized Leader) algorithm.

[0072] Among them, the fault feature extraction model online optimizer is used to update the weight parameters of the conditional variational autoencoder-gated recurrent unit network model.

[0073] Step 206, obtain fault information and input it into the CVAE-GRU model after updating the weight parameters, obtain the fault features output by the CVAE-GRU model and calculate the anomaly score based on the fault features, and output the comparison result between the anomaly score and the preset threshold.

[0074] Among them, this application selects the entropy weight method to dynamically calculate the weight of each dimension feature, and then the weighted sum is used to obtain the anomaly score. Assuming the anomaly score is S, the calculation formula is as follows:

[0075]

[0076] Among them, W i is the weight of the i-th dimension, F i is the eigenvalue of the i-th dimension, E i is the information entropy of the i-th dimension, and k is the fault feature dimension.

[0077] In the above-mentioned spacecraft attitude control system anomaly detection method, a feature extraction model is obtained by constructing a CVAE-GRU network; and a model optimizer is designed to update the model weight parameters online; then, the streaming data peak algorithm is used to calculate and update the adaptive threshold in real time to achieve real-time dynamic detection of anomalies.

[0078] In one embodiment, the CVAE network structure is as follows: Figure 3 As shown in the figure, CVAE introduces additional label information c on the basis of VAE. The network itself only encodes and randomly samples x, and the label information does not need to pass through the encoding layer. It is directly input into the hidden layer after one-hot encoding conversion, and directly input into the decoding layer after being spliced ​​with the latent space variable z. This network structure not only reduces the dimensionality of the complex multi-dimensional information of the reconstructed input x, but also retains the one-dimensional discrete label information of the label c, making the reconstruction result more accurate, and the output category can be adjusted by the input label information. Its loss function is as follows:

[0079]

[0080] It is worth mentioning that since it is difficult to directly calculate μ and σ based on the input x in the encoding layer, two fully connected layers are used to fit μ and σ respectively; on the other hand, the loss function is Generating samples is a sampling process, and the gradient of back propagation cannot be calculated, so the re-parameterization technique needs to be introduced.

[0081] The specific method is to sample ε from a standard normal distribution N(0,1) and then calculate z. The specific calculation process is as follows:

[0082] z=μ+∈×σ;

[0083] In one embodiment, the GRU network structure and its unit internal structure are as follows: Figure 4 and Figure 5 As shown in the figure, compared with other prediction models, GRU has a simplified internal neuron structure, reduces the number of model parameters, reduces the network training cycle, and improves the calculation speed of the model itself, reducing the occupation of computing resources while ensuring the prediction accuracy. The specific calculation process includes:

[0084] z t =σ(w z x t +u z h t-1 );

[0085] r t =σ(w r x t +u r h t-1 );

[0086]

[0087]

[0088] y t =σ(w y h t );

[0089] Among them, r t is the reset gate output, z t is the update gate output, σ is the sigmoid activation function, is the tanh activation function, x t Input for the current moment, is the candidate hidden layer state at the current moment, h t is the hidden layer state at the current moment, h t-1 is the hidden layer state at the previous moment.

[0090] In one embodiment, according to the model functions and input and output parameters, the actuator, dynamics and kinematics models are designed respectively, and the feature extraction network is trained in groups;

[0091] The reconstruction input parameter with a time sliding window length of t is selected, and the label corresponding to the mode of the maneuvering mode flag in the window is used as the window label c. Three CVAE network reconstruction models are established: CVAE-1 network reconstructs the flywheel motor current i(1:t) and flywheel control torque T at the previous t moments. c(1:t); CVAE-2 network reconstructs the flywheel speed n(1:t) at the previous t moments; CVAE-3 network reconstructs the angular velocity ω(1:t) at the previous t moments. If the value of t is too small, the reconstruction effect will be poor, while if t is too large, the real-time performance of the algorithm will deteriorate. After comprehensively considering the reconstruction effect and real-time factors, the reconstruction window length t is selected as 20.

[0092] A prediction model combining three GRU networks and a multi-head attention mechanism is established and the input parameters with a time sliding window length of t are selected: the GRU-1 network uses the flywheel motor current i(1:t) at the previous t moments, the control torque T c (1:t), flywheel speed n(1:t) as input, predict the speed at time t+1 The GRU-2 network uses the flywheel speed n(1:t) and attitude angular velocity ω(1:t) of the previous t moments as input to predict the angular velocity at time t+1 The GRU-3 network uses the angular velocity ω(1:t) and attitude angle of the previous t moments As input, predict the attitude angle at time t+1 In order to keep consistent with the output length of the CVAE network, the prediction input window length t is selected as 20 and the output step size is 1.

[0093] like Figure 8-Figure 11 As shown, the reconstruction effect of CVAE is Figure 12-14 As shown in the figure, the prediction effect of GRU; the reconstruction accuracy of CVAE is measured by the mean square error MSE and R2 scores between the reconstruction value of CVAE and the true value. The calculation formulas of MSE and R2 scores are as follows:

[0094]

[0095] The reconstruction accuracy calculation results are shown in Table 1:

[0096] Parameter name MSE R2 score Parameter name MSE R2 score Flywheel 1 current <![CDATA[4.263×10 -4 ]]> 0.9786 Flywheel 1 speed 6.685 0.9953 Flywheel 2 current <![CDATA[3.472×10 -4 ]]> 0.9806 Flywheel 2 speed 5.979 0.9959 Flywheel 3 current <![CDATA[4.191×10 -4 ]]> 0.9781 Flywheel 3 speed 7.005 0.9959 Flywheel 4 current <![CDATA[3.949×10 -4 ]]> 0.9804 Flywheel 4 speed 6.679 0.9958 X-axis control torque <![CDATA[7.009×10 -5 ]]> 0.9709 Roll angular velocity <![CDATA[6.124×10 -7 ]]> 0.9815 Y-axis control torque <![CDATA[8.937×10 -5 ]]> 0.9735 Pitch angular velocity <![CDATA[6.937×10 -7 ]]> 0.9840 Z-axis control torque <![CDATA[4.551×10 -5 ]]> 0.9870 Yaw angular velocity <![CDATA[2.11×10 -10 ]]> 0.9826

[0097] Table 1;

[0098] The prediction accuracy of GRU is measured using the mean square error MSE and R2 score between the predicted value of GRU and the true value. The results are shown in Table 2:

[0099] Parameter name MSE R2 score Parameter name MSE R2 score Flywheel 1 speed 9.727 0.9932 Pitch angular velocity <![CDATA[9.779×10 -7 ]]> 0.9774 Flywheel 2 speed 9.219 0.9937 Yaw angular velocity <![CDATA[2.744×10 -10 ]]> 0.9785 Flywheel 3 speed 9.912 0.9943 Roll Angle <![CDATA[1.630×10 -3 ]]> 0.9992 Flywheel 4 speed 9.172 0.9943 Pitch Angle <![CDATA[2.388×10 -4 ]]> 0.9994 Roll angular velocity <![CDATA[7.461×10 -7 ]]> 0.9734 Yaw angle <![CDATA[9.650×10 -10 ]]> 0.9993

[0100] Table 2;

[0101] Based on the CVAE-GRU fault feature extraction model, the FTRL algorithm is used to design an online optimizer for the fault feature extraction model, and the model weight parameters are updated online to improve the adaptability of the feature extraction function to on-orbit data.

[0102] The updated model is used to output fault features and perform exponential weighted smoothing on them. Then, the entropy weight method is used to calculate the anomaly score. Finally, the streaming data peak algorithm is used to calculate and update the adaptive threshold in real time. By comparing the anomaly score and threshold, real-time dynamic detection of anomalies is achieved.

[0103] The fault feature extraction scheme based on the conditional variational autoencoder-gated recurrent unit network (CVAE-GRU) model is as follows: Fig.15 As shown, including:

[0104] Data preprocessing: including outlier removal, resampling and missing value filling.

[0105] The hard threshold method is used to roughly remove outliers. According to the physical characteristics of the telemetry parameters, appropriate upper and lower limits of outliers are selected to remove outliers from the data, and the removed points are replaced with null values. Fig.16 As shown in the figure, for the three-axis attitude angle, the normal value range is [-180°, 180°], so -180 is set as the lower threshold and 180 is set as the upper threshold, and the outliers of the attitude angle telemetry data are eliminated. Fig.16 This is before outliers are removed (left picture). Fig.16 After outliers are removed (right picture);

[0106] After removing outliers, sample the data with different sampling times, retain the median in each fixed sampling interval as the downsampling result value of the interval, and align the time axis of each parameter, such as Fig.17 As shown, Fig.17 (Left) is before resampling. Fig.17 (Right) after resampling;

[0107] For the null value data points after outlier removal and upsampling, the forward interpolation method is used to fill in the missing values, such as Fig.18 As shown, Fig.18 (Left picture) is before missing value completion. Fig.18 (Right picture) shows the missing values ​​after completion.

[0108] Ground model training: The telemetry parameters are divided into three groups, and the reconstruction model CVAE and the prediction model GRU are trained using normal data to fit the flywheel, dynamics and kinematics models. The first group of networks is responsible for fitting the flywheel model, using the CVAE-1 model to fit the flywheel motor current i and the flywheel control torque T cThe second group of networks is responsible for fitting the dynamic model, reconstructing the flywheel speed n using the CVAE-2 model, and predicting the angular velocity ω using the GRU-2 model; the third group of networks is responsible for fitting the kinematic model, reconstructing the angular velocity ω using the CVAE-3 model, and predicting the attitude angle using the GRU-3 model.

[0109] On-orbit operation: Use the three groups of CVAE and GRU models trained on the ground, use the CVAE reconstruction output of each group as the input of GRU, and combine them to form three groups of CVAE-GRU models to reconstruct and predict the on-orbit data in real time, and obtain the predicted values ​​of flywheel speed, attitude angular velocity, and attitude angle, and calculate the final fault characteristics with the sensor measurement values.

[0110] In order to prove the advantage of combining reconstruction and prediction in the feature extraction method studied, this experiment removes the reconstruction part and the prediction part respectively, and then conducts ablation experiments with the complete algorithm. The feature extraction model without the CVAE part is named model A, and the feature extraction model without the GRU and multi-head attention mechanism part is named model B. The entropy weight method is used uniformly to reduce the dimension of the residual features generated by the model to form anomaly scores. The anomaly score ROC curves of the three models on the real satellite telemetry dataset are shown in the figure. Fig.19 As shown, the AUROC values ​​calculated according to the curve are shown in Table 3.

[0111] Model Model A Model B CVAE-GRU AUROC 0.9697 0.9106 0.9719

[0112] Table 3;

[0113] The anomaly score obtained by the CVAE-GRU-based feature extraction method proposed in this application has the largest AUROC area on both the telemetry data and the simulation data data sets, which verifies that the method combining CVAE and GRU has a good feature extraction effect.

[0114] For the online optimizer of the fault feature extraction model designed based on the FTRL algorithm, the weight update formula is as follows:

[0115]

[0116] in G (1:t) is the cumulative gradient from step 1 to step t, λ1 and λ2 are hyperparameters, corresponding to the influence of L1 and L2 regularization respectively. The weight update formula of step t+1 of the i-th dimension W is further derived as follows:

[0117]

[0118] Each dimension of FTRL features is updated at a different learning rate. This improvement takes into account the uneven distribution of online input samples in different dimensions. If a feature changes quickly, its corresponding learning rate will also decrease quickly. The learning rate of the feature at the tth step of the i-th dimension is:

[0119]

[0120] Among them, α and β are hyperparameters that can control the rate of change of the learning rate. is the gradient of the s-th step in the i-th dimension.

[0121] A comparative experiment was conducted using the anomaly detection method with and without the FTRL feature extraction optimizer. A set of anomaly detection result data was selected as a typical case, such as Fig. 20 As shown, Fig. 20 (Left) does not use the FTRL optimizer. Fig. 20 (Right) shows the use of FTRL feature optimizer. The threshold part uses adaptive threshold for detection. The data of the first 5000 seconds are selected to initialize the adaptive threshold. The parameter data begins to change dynamically around 5800 seconds, and the anomaly occurs around 17300 seconds. As can be seen from the figure, the anomaly score without FTRL feature extraction optimizer begins to increase significantly after 5800 seconds. In contrast, the anomaly score with FTRL feature extraction optimizer is distributed consistently in all normal intervals, which shows that the FTRL optimizer can make the feature extraction model adapt to the dynamic changes of on-orbit parameters.

[0122] from Fig. 20 As can be seen from the enlarged figure in , the adaptive threshold maintains a certain update frequency when the data changes dynamically by screening the extreme points of the data, and tends to a more stable value in the abnormal segment. In the case without the FTRL feature extraction optimizer, the data changes of the anomaly score in the normal segment are more drastic, and the adaptive threshold is difficult to adapt to such frequent changes, resulting in many false detections. In the case with the FTRL feature extraction optimizer, the data changes of the anomaly score in the normal segment are relatively stable, and the adaptive threshold can also be dynamically adjusted according to the extreme value distribution to achieve a good anomaly detection effect.

[0123] Evaluation indicators Adaptive Threshold FTRL+Adaptive Threshold Accuracy 90.70％ 100％ Recall 99.67％ 99.70％ F1 score 94.97％ 99.85％ False alarm rate 5.10％ 0％

[0124] Table 4;

[0125] As shown in Table 4, in the satellite telemetry dataset, the anomaly detection method using the FTRL feature extraction optimizer has a 9.30% higher precision, a 0.03% higher recall, a 4.88% higher F1 score, and a 5.10% lower false alarm rate than the method without the FTRL optimizer. The method without the FTRL feature extraction optimizer has a lower precision and a higher false alarm rate, indicating that it often detects normal data as anomalies, which is also consistent with Fig. 20 The fact that there are many false detection phenomena.

[0126] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0127] Based on the same inventive concept, the embodiment of the present application also provides a computer device for detecting anomalies in a spacecraft attitude control system for implementing the above-mentioned method for detecting anomalies in a spacecraft attitude control system. The implementation scheme for solving the problem provided by the computer device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the computer device for detecting anomalies in a spacecraft attitude control system provided below can refer to the limitations of the method for detecting anomalies in a spacecraft attitude control system above, and will not be repeated here.

[0128] In one embodiment, a computer device for detecting anomalies of a spacecraft attitude control system is provided, comprising: a constructed module, a model online optimization module, and a fault detection module, wherein:

[0129] The constructed module, namely the conditional variational autoencoder-gated recurrent unit network model;

[0130] Model online optimization module, used to update the weight parameters of the conditional variational autoencoder-gated recurrent unit network model in real time online;

[0131] The fault detection module is used to obtain telemetry parameters and input them into the conditional variational autoencoder-gated recurrent unit network model after updating the weight parameters, obtain the fault features output by the conditional variational autoencoder-gated recurrent unit network model and calculate the anomaly score based on the fault features, and output the comparison result between the anomaly score and the preset threshold.

[0132] In one embodiment, the constructed module is also used to: select a corresponding standard threshold range based on the physical characteristics of the telemetry parameters, and eliminate outliers in the telemetry parameters based on the standard threshold range; use the telemetry parameters after eliminating outliers to train a reconstruction model and a prediction model, respectively, the reconstruction model is a conditional variational autoencoder, and the prediction model is a gated recurrent unit network; reconstruct and predict on-orbit data based on the reconstruction model and the prediction model and obtain corresponding prediction values; and calculate and obtain fault characteristics based on the prediction values ​​and corresponding measurement values.

[0133] In one embodiment, the model online optimization module is also used to: obtain telemetry parameters of a sliding window of preset length; set the label corresponding to the majority of the maneuvering mode flag in the window as the window label; respectively construct an actuator model, a dynamic model and a kinematic model based on the conditional variational autoencoder-gated recurrent unit network model; and respectively train the actuator model, the dynamic model and the kinematic model based on the telemetry parameters and the window label.

[0134] In one embodiment, the model online optimization module is also used to: construct a prediction model and input telemetry parameters of a time sliding window of preset length into the prediction model; predict the flywheel speed at a future time based on the historical flywheel motor current, historical control torque and historical flywheel speed in the historical telemetry parameters; predict the flywheel angular velocity at a future time based on the historical flywheel speed and historical attitude angular velocity in the historical telemetry parameters; predict the flywheel attitude angle at a future time based on the historical angular velocity and historical attitude angle in the historical telemetry parameters.

[0135] In one embodiment, the model online optimization module is also used to: perform verification operations based on the reconstructed values ​​and true values ​​corresponding to the reconstructed model and the prediction model: calculate the mean square error based on the reconstructed values ​​and the true values; calculate the goodness of fit based on the reconstructed values ​​and the true values.

[0136] In one embodiment, the fault detection module is also used to: obtain the loss function corresponding to the conditional variational autoencoder-gated recurrent unit network model; calculate the cumulative gradient of the loss function of a preset round and obtain the model parameter update formula; calculate the update value of the parameters of each dimension based on the optimizer hyperparameters and the cumulative gradient of the parameters.

[0137] Each module in the above-mentioned spacecraft attitude control system abnormality detection device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0138] Those skilled in the art will understand that Fig.21 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0139] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0140] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0141] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0142] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0143] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for detecting anomalies in a spacecraft attitude control system, characterized in that: The method comprises: Construct a conditional variational autoencoder-gated recurrent unit network model; For the conditional variational autoencoder-gated recurrent unit network model, constructing a fault feature extraction model online optimizer, wherein the fault feature extraction model online optimizer is used to update weight parameters of the conditional variational autoencoder-gated recurrent unit network model; Obtain telemetry parameters and input them into the conditional variational autoencoder-gated recurrent unit network model after updating weight parameters, obtain fault features output by the conditional variational autoencoder-gated recurrent unit network model and calculate anomaly scores based on the fault features, and output a comparison result between the anomaly scores and a preset threshold; The output of the conditional variational autoencoder in the conditional variational autoencoder-gated recurrent unit network model is used as the input of the gated recurrent unit network.

2. The method according to claim 1, characterized in that The construction of the conditional variational autoencoder-gated recurrent unit network model includes: Selecting a corresponding standard threshold range based on the physical characteristics of the telemetry parameters, and eliminating outliers of the telemetry parameters based on the standard threshold range; The telemetry parameters after outliers are removed are used to train a reconstruction model and a prediction model respectively, wherein the reconstruction model is a conditional variational autoencoder and the prediction model is a gated recurrent unit network; Reconstruct and predict the on-orbit data based on the reconstruction model and the prediction model and obtain the corresponding prediction value; The fault characteristics are calculated based on the predicted values ​​and the corresponding measured values.

3. The method according to claim 2, characterized in that The telemetry parameters after outliers are removed are used to train the reconstruction model including: Get telemetry parameters of a sliding window of preset length; The label corresponding to the majority of the motor mode flag bits in the window is set as the window label; The actuator model, dynamics model and kinematics model based on the conditional variational autoencoder-gated recurrent unit network model are constructed respectively; The actuator model, dynamic model and kinematic model are trained based on telemetry parameters and window labels respectively.

4. The method according to claim 2, characterized in that: The telemetry parameters after outliers are removed are used to train the prediction model, including: constructing a prediction model and inputting telemetry parameters of a time sliding window of a preset length into the prediction model; Predicting the flywheel speed at a future time based on the historical flywheel motor current, historical control torque and historical flywheel speed in the historical telemetry parameters; Predict the flywheel angular velocity at the future time based on the historical flywheel speed and historical attitude angular velocity in the historical telemetry parameters; The flywheel attitude angle at the future moment is predicted based on the historical angular velocity and historical attitude angle in the historical telemetry parameters.

5. The method according to claim 4, characterized in that The method further comprises: Verification operation is performed based on the reconstructed values ​​and true values ​​corresponding to the reconstructed model and the predicted model: Calculate the mean square error based on the reconstructed value and the true value; Calculate the goodness of fit based on the reconstructed and true values.

6. The method according to claim 1, characterized in that For the conditional variational autoencoder-gated recurrent unit network model, constructing an online optimizer for a fault feature extraction model includes: Get the loss function corresponding to the conditional variational autoencoder-gated recurrent unit network model; Calculate the cumulative gradient of the loss function of the preset rounds and the corresponding regularized hyperparameters to obtain the weight update formula; The updated values ​​of the parameters of each dimension are calculated based on the optimizer hyperparameters and parameter cumulative gradients.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having stored thereon a computer program and parameters of the model according to claim 1, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images