A satellite fault detection method based on spatiotemporal dynamic feature fusion

Abstract

The application discloses a satellite fault detection method based on space-time dynamic feature fusion, and comprises the following steps: S1, a self-encoder neural network based on a bidirectional LSTM algorithm is designed, and is denoted as TSAD TFN ; a time dynamic feature variable of data is extracted from input satellite data, and is denoted as Latent TFN ; S2, a self-encoder neural network based on a convolutional self-encoder is designed, and is denoted as TSAD SFN ; a space dynamic feature variable of data is extracted from input satellite data, and is denoted as Latent SFN ; S3, a space-time dynamic feature variable of satellite data is obtained by fusing the time dynamic feature variable and the space dynamic feature variable generated in steps S1 and S2, and is denoted as Latent TSFN In the application, time dependence and space dependence in satellite data are simultaneously captured, deep-level change rules of satellite data are effectively mined, and satellite data features are extracted, so that the satellite data features can be used not only in fault detection but also in technical fields such as data prediction, and satellite fault detection without a fault label is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite fault detection, and particularly relates to a satellite fault detection method based on spatiotemporal dynamic feature fusion. BACKGROUND

[0002] At present, satellite fault detection methods include parameter threshold method, expert system method and data-driven method, and have been widely researched and applied. Although scholars at home and abroad have made relevant achievements in satellite fault detection, due to the characteristics of numerous satellite parameters, complex spatiotemporal relationship and strong professionalism, the existing satellite fault detection methods still have deficiencies, and new satellite fault detection methods need to be researched.

[0003] The most widely used engineering application in the field of satellite fault detection is the parameter threshold upper and lower limit judgment and the expert system. The former detects faults by judging whether the telemetry parameters exceed the set upper and lower limit range, and the latter detects faults by using expert experience to summarize typical fault modes and using rule matching. These two methods have clear rules and strong interpretability, but they are heavily dependent on expert knowledge and cannot detect abnormal events outside prior knowledge. In addition, satellite data includes multiple dimensional parameters such as temperature and pressure, and there are obvious changes between different dimensional parameters at the fault moment. It is difficult to model the relationship between parameters by relying on expert systems and threshold methods only. Therefore, data-driven fault detection methods have gradually attracted attention. By utilizing the large amount of data generated by satellites, faults can be detected, which can effectively overcome the shortcomings of threshold upper and lower limits and expert system methods that rely on expert prior knowledge. For data-driven methods, how to effectively mine the change rules of satellite data is the key to determining their fault detection performance. However, due to the rarity of satellite fault conditions, the number of positive and negative samples in the data is extremely unbalanced. In addition, due to the professionalism and complexity of satellite mechanisms, it is difficult to cover all types of fault samples, resulting in a lack of labeled training data. Therefore, existing data-driven methods rely on fault labels and cannot effectively mine satellite data rules, making it difficult to effectively implement satellite fault detection. Unsupervised methods such as PCA and clustering are also limited in detecting high-dimensional and dynamically changing data generated by satellites.

[0004] Therefore, the present application proposes a satellite fault detection method based on spatiotemporal dynamic feature fusion. SUMMARY

[0005] The purpose of the present application is to solve the above problems and propose a satellite fault detection method based on spatiotemporal dynamic feature fusion.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A satellite fault detection method based on spatiotemporal dynamic feature fusion, comprising the following steps:

[0008] S1. Design an autoencoder neural network based on bidirectional LSTM algorithm, denoted as TSAD TFN , extract the temporal dynamic feature variable of the data from the input satellite data, denoted as Latent TFN ;

[0009] S2. Design an autoencoder neural network based on convolutional autoencoder, denoted as TSAD SFN , extract the spatial dynamic feature variable of the data from the input satellite data, denoted as Latent SFN ;

[0010] S3. According to the temporal dynamic feature variable and the spatial dynamic feature variable generated in steps S1 and S2, the spatio-temporal dynamic feature variable of the satellite data is fused, denoted as Latent TSFN ;

[0011] S4. Design a neural network based on Transformer, denoted as TSAD Predict , TSAD Predict , according to the spatio-temporal dynamic feature variable generated in step S3, predict the value of the telemetry parameter after P frames;

[0012] S5. Use the telemetry data generated by the satellite in normal working condition to train the neural network model composed of steps S1-S4 end to end, obtain the trained model, denoted as TSAD;

[0013] S6. Input all samples in the training data set into the model generated in step S5, calculate the fault score of all samples in the training set;

[0014] S7. According to the fault score set generated in step S6 on the training data set, calculate the satellite fault detection threshold value by using the POT method;

[0015] S8. Input the test set data into the model TSAD, sequentially go through steps S1-S7, obtain the detection results of all test set samples, calculate the performance indicators P, R, F1 of the algorithm on the test set, and repeat steps S1 to S8 to optimize the performance of the method;

[0016] S9. Input the satellite data sample to be detected into the model, generate the final prediction residual from step S8, and obtain the final detection result according to the threshold value generated in step S7.

[0017] Preferably, the step S1 extracts the satellite data temporal dynamic feature, which specifically includes the following steps:

[0018] S11. Process the original telemetry data by using a sliding window to obtain a single input sample of M-dimensional continuous K frames, and the formula is as follows:

[0019]

[0020]

[0021] where S input represents a single input sample; X i represents the M-dimensional data of the i-th frame, 1≤i≤K;

[0022] S12. Design a bidirectional LSTM-based autoencoder neural network TSAD TFN , including an encoder network and a decoder network, respectively denoted as TFN Encoder and TFN decoder .

[0023] S13. The encoder network TFN Encoder extracts the time feature encoding from the output S input of the step, and the formula is as follows:

[0024] Latent TFN = TFN encoder (S input )

[0025] where S input represents a single input sample; Latent TFN represents the time feature encoding;

[0026] S14. The decoder network decodes the time feature encoding extracted in step, and the formula is as follows:

[0027] O TFN = TFN decoder (Latent TFN )

[0028] where O TFN represents the data output by the decoder network;

[0029] S15. Calculate the reconstruction error of the input S TFN and the output O input of the autoencoder neural network TSAD TFN , which is part of the loss function in step, to complete the training of the model, and the formula is as follows:

[0030]

[0031] Preferably, the method of extracting spatial dynamic features of satellite data in the step S2 specifically comprises the following steps:

[0032] S21. Design a convolution-based autoencoder network TSAD SFN , including an encoder network and a decoder network, respectively denoted as SFNEncoder and SFN decoder ;

[0033] S22. Encoder network SFN Encoder extracts spatial feature encoding from the output S input of step S11, as follows:

[0034] Latent SFN = SFN encoder (S input )

[0035] wherein: Latent SFN represents spatial feature encoding;

[0036] S23. Decoder network decodes the spatial feature encoding extracted in step S23, as follows:

[0037] O SFN = SFN decoder (Latent SFN )

[0038] wherein: O SFN represents data output by the decoder network;

[0039] S24. The reconstruction error of the input S SFN and the output O input of the autoencoder network TSAD SFN is calculated in step S5 as part of the loss function, completing the training of the model, as follows:

[0040]

[0041] Preferably, the step S3 fuses the spatiotemporal dynamic features of the satellite data, as follows:

[0042] Latent TSFN = [Latent TFN , Latent SFN ]

[0043] wherein: Latent TSFN represents the extracted spatiotemporal dynamic features of the satellite data sample.

[0044] Preferably, the step S4 predicts the value of the satellite data after P frames, specifically:

[0045] S41. Designing a satellite data prediction network TSAD Predict based on Transformer;

[0046] S42. The spatiotemporal dynamic features LatentTSFN TSAD (Tracking Satellite Data Prediction Network) Predict The predicted value is obtained using the following formula:

[0047] O Predict =TSAD Predict (Latent TSFN )

[0048] In the formula: O Predict This represents the predicted data after P frames.

[0049] S43. Calculate the predicted network output O Predict The distance from the actual value after P frames is calculated using the following formula:

[0050]

[0051] In the formula: Y P This represents the actual data after the P-frame.

[0052] Preferably, the method for training the neural network model TSAD in step S5 specifically includes the following steps:

[0053] S51. The neural network TSAD consists of three subnetworks, namely TSAD TFN TSAD SFN and TSAD Predict ;

[0054] S52. Select M analog telemetry parameters based on the satellite equipment to be fault detected;

[0055] S53. Select L consecutive frames of data for the M telemetry parameters under normal operating conditions of the satellite as the training set, denoted as X. train The formula is as follows:

[0056]

[0057] In the formula: L represents the number of frames in the training set data, X i ,1≤i≤L represents the M-dimensional data of the i-th frame;

[0058] S54. Use the "spectral residual" method to remove training set X train The wild value in the middle, to obtain the new X train ;

[0059] S55. Calculate the training set X train mean μ train and standard deviation δ train , for X train All samples in the dataset are normalized using the following formula:

[0060]

[0061] S56. Set the sliding window width as K and the sliding step as S, and slide the training set X train to obtain a new training set X train ;

[0062] S57. Repeat step S56 until the loss Loss train of the model TSAD is minimized, and output the validation set X valid corresponding to the training set X valid , wherein the number of samples of the validation set X train is 20% to 30% of the number of samples in X

[0063] S58. Calculate the loss of the model TSAD according to the loss of the neural networks TSAD TFN , TSAD SFN and TSAD Predict obtained in steps S15, S24 and S43, and the formula is as follows:

[0064] Loss TSAD = λ1Loss TFN + λ2Loss SFN + λ3Loss Predict

[0065] wherein Loss TSAD represents the loss of the neural network TSAD, λ1, λ2 and λ3 are hyperparameters, and respectively represent the weight proportion of the neural networks TSAD TFN , TSAD SFN and TSAD Predict in the total loss Loss TSAD ;

[0066] S59. Train the model TSAD by using the back propagation algorithm to minimize the loss Loss TSAD of step S56.

[0067] Preferably, the method of step S6 for calculating the failure score of all samples in the training set comprises the following steps:

[0068] S61. Take the L2 distance between the output O Predict of the prediction network and the actual value of the parameter of the P frame as the failure score of a single sample;

[0069] S62. Traverse all data samples in the training set, and calculate the set of failure scores of the model on the training set.

[0070] Preferably, the calculation formula of step S8 is as follows:

[0071]

[0072]

[0073]

[0074] In the formula, TP represents the number of samples correctly classified as faults, FP represents the number of samples incorrectly classified as faults, TN represents the number of samples correctly classified as normal, and FN represents the number of samples incorrectly classified as normal.

[0075] In summary, due to the adoption of the technical solutions described above, the beneficial effects of the present application are:

[0076] In the present application, by simultaneously capturing the time dependence and spatial dependence in satellite data, the deep-level change rule of satellite data is effectively mined, and the extracted satellite data features can be used not only for fault detection but also for data prediction and other technical fields, realizing satellite fault detection without fault labels. BRIEF DESCRIPTION OF DRAWINGS

[0077] Fig. 1 Fig. 1 shows a general block diagram structure schematic diagram of a satellite fault detection method based on spatiotemporal dynamic feature fusion according to an embodiment of the present application;

[0078] Fig. 2 Fig. 1 shows a general block diagram structure schematic diagram of a satellite fault detection method based on spatiotemporal dynamic feature fusion according to an embodiment of the present application; DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0080] Please refer to Figs. 1-2 The present application provides a technical solution:

[0081] 1. A satellite fault detection method based on spatiotemporal dynamic feature fusion, comprising the following steps:

[0082] S1. Design a self-encoder neural network based on a bidirectional LSTM algorithm, denoted as TSAD TFN Extract the time dynamic feature variable of the data from the input satellite data, denoted as Latent TFN ;

[0083] S2. Design a self-encoder neural network based on a convolutional self-encoder, denoted as TSADSFN , from the input satellite data, denoted as Latent SFN ;

[0084] S3. According to the time dynamic feature variables and the space dynamic feature variables generated in steps S1 and S2, spatio-temporal dynamic feature variables of the satellite data are fused, denoted as Latent TSFN ;

[0085] S4. A neural network based on Transformer is designed, denoted as TSAD Predict , TSAD Predict According to the spatio-temporal dynamic feature variables generated in steps, the value of the telemetry parameter after P frames is predicted.

[0086] S5. The neural network model composed of steps S1-S4 is trained end-to-end using the telemetry data generated by the satellite under normal working conditions to obtain a trained model, denoted as TSAD.

[0087] S6. All samples in the training data set are input into the model generated in step S5 to calculate the fault scores of all samples in the training set.

[0088] S7. According to the set of fault scores generated in step S6 on the training data set, the POT method is used to calculate the satellite fault detection threshold.

[0089] S8. The test set data is input into the model TSAD, and sequentially undergoes steps S1-S7 to obtain the detection results of all test set samples, calculates the performance indicators P, R, F1 of the algorithm on the test set, and repeats steps S1 to S8 to optimize the performance of the method.

[0090] S9. The satellite data sample to be detected is input into the model to generate the final prediction residual, and according to the threshold generated in step S7, the final detection result is obtained.

[0091] Specifically, as shown in FIG. 1, the method for extracting the time dynamic feature of the satellite data in step S1 specifically includes the following steps: Fig. 2

[0092] S11. The original telemetry data is processed using a sliding window to obtain a single input sample of M-dimensional continuous K frames, and the formula is as follows:

[0093]

[0094]

[0095] In the formula, S input represents a single input sample; X i ,1≤i≤K represents the M-dimensional data of the i-th frame.​

[0096] S12. Design a bidirectional LSTM-based autoencoder neural network TSAD TFN It consists of two parts: an encoder network and a decoder network, denoted as TFN respectively. Encoder and TFN decoder ;

[0097] S13. Encoder Network TFN Encoder From the output S of the step input The time feature encoding is extracted using the following formula:

[0098] Latent TFN =TFN encoder (S input )

[0099] In the formula: S input Represents a single input sample; Latent TFN Represents time feature encoding;

[0100] S14. The decoder network decodes the temporal feature encoding extracted in the step, as shown in the following formula:

[0101] O TFN =TFN decoder (Latent TFN )

[0102] In the formula: O TFN This represents the data output by the decoder network;

[0103] S15. Calculate the autoencoder neural network TSAD TFN Input S input With output O TFN The reconstruction error is incorporated as part of the loss function in this step to complete the model training, as shown in the following formula:

[0104]

[0105] Specifically, such as Fig. 2 As shown, the method for extracting spatial dynamic features of satellite data in step S2 specifically includes the following steps:

[0106] S21. Design a convolution-based autoencoder network TSAD SFN It consists of two parts: an encoder network and a decoder network, denoted as SFN respectively. Encoder and SFN decoder ;

[0107] S22. Encoder Network SFN Encoder From the output S of step S11 input The spatial feature encoding is extracted from it, and the formula is as follows:

[0108] Latent SFN =SFN encoder (S input )

[0109] In the formula: Latent SFN Represents spatial feature encoding;

[0110] S23. The decoder network decodes the spatial feature encoding extracted in step S23, using the following formula:

[0111] O SFN =SFN decoder (Latent SFN )

[0112] In the formula: O SFN This represents the data output by the decoder network;

[0113] S24. Calculate the autoencoder network TSAD SFN Input S input With output O SFN The reconstruction error is incorporated as part of the loss function in step S5 to complete the model training, as shown in the following formula:

[0114]

[0115] Specifically, such as Fig. 2 As shown, step S3 fuses the spatiotemporal dynamic features of the satellite data, and the formula is as follows:

[0116] Latent TSFN =[Latent TFN Latent SFN ]

[0117] In the formula: Latent TSFN This indicates the spatiotemporal dynamic characteristics of the extracted satellite data samples.

[0118] Specifically, such as Fig. 2 As shown, step S4 predicts the value of the satellite data after frame P, specifically as follows:

[0119] S41. Design a Transformer-based satellite data prediction network TSAD Predict ;

[0120] S42. The spatiotemporal dynamic features obtained in step S3 are Latent TSFN TSAD (Tracking Satellite Data Prediction Network) Predict The predicted value is obtained using the following formula:

[0121] O Predict= TSAD Predict (Latent TSFN )

[0122] wherein O Predict represents the predicted post-P frame data.

[0123] S43. Calculate the distance between the predicted network output O Predict and the actual value of the post-P frame, as follows:

[0124]

[0125] wherein Y P represents the actual post-P frame data.

[0126] Specifically, as shown in FIG. 5, the method for training the neural network model TSAD includes the following steps: Fig. 2

[0127] S51. The neural network TSAD includes three sub-networks, namely TSAD TFN , TSAD SFN and TSAD Predict .

[0128] S52. Select M analog telemetry parameters according to the satellite equipment to be detected for faults;

[0129] S53. Select the data of L consecutive frames of the M telemetry parameters of the satellite under normal operating conditions as the training set, denoted as X train , as follows:

[0130]

[0131] wherein L represents the number of frames of the training set data, X i ,1≤i≤L represents the M-dimensional data of the ith frame;

[0132] S54. Remove outliers in the training set X train using the "spectral residual" method to obtain a new X train ;

[0133] S55. Calculate the mean μ train and standard deviation δ train of the training set X train , and normalize all samples in X train , as follows:

[0134]

[0135] S56. Set the sliding window width to K and the sliding step to S, and calculate the distance between the predicted network output O train ​The new training set X is obtained by continuously sliding up. train ;

[0136] S57. Repeat steps until outputting the training set X. train The corresponding validation set X valid , verification set X valid The number of samples is X train 20% to 30% of the sample size;

[0137] S58. Based on the neural network TSAD obtained in steps S15, S24, and S43 TFN TSAD SFN and TSAD Predict The loss of the TSAD model is calculated using the following formula:

[0138] Loss TSAD =λ1Loss TFN +λ2Loss SFN +λ3Loss Predict

[0139] In the formula: Loss TSAD Let λ1, λ2, and λ3 represent the loss of the neural network TSAD, and λ3 be hyperparameters representing the loss of the neural network TSAD. TFN TSAD SFN and TSAD Predict In total loss TSAD The weighting percentage of the content;

[0140] S59. Minimize the loss in step S56. TSAD With the goal of training the model TSAD using the backpropagation algorithm.

[0141] Step S6, calculating the fault scores for all samples in the training set, specifically includes the following steps:

[0142] S61. To predict the network output O Predict The L2 distance between the actual parameter value after P frames is used as the fault score for a single sample.

[0143] S62. Traverse all data samples in the training set and calculate the fault score set of the model on the training set.

[0144] Specifically, such as Fig. 2 As shown, step S8 uses the following calculation formula:

[0145]

[0146]

[0147]

[0148] wherein TP represents the number of samples correctly classified as faulty, FP represents the number of samples incorrectly classified as faulty, TN represents the number of samples correctly classified as normal, and FN represents the number of samples incorrectly classified as normal

[0149] In summary, the satellite fault detection method based on spatio-temporal dynamic feature fusion provided in the embodiment is applied in communication / navigation satellite telemetry data analysis and fault detection. By training a model using satellite historical data, detection of faults in load, control, TT&C, power supply and distribution subsystems is achieved, and the dependence on a fault knowledge base is eliminated.

[0150] The above description of the embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Accordingly, the present application is not to be limited to the embodiments shown herein but is to be accorded the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A satellite fault detection method based on spatiotemporal dynamic feature fusion, characterized in that, Includes the following steps: S1. Design an autoencoder neural network based on the bidirectional LSTM algorithm, denoted as TSAD. TFN Extract the time-dynamic feature variables of the input satellite data, denoted as Latent. TFN ; S2. Design an autoencoder neural network based on a convolutional autoencoder, denoted as TSAD. SFN The spatial dynamic characteristic variables of the data are extracted from the input satellite data, denoted as Latent. SFN ; S3. Based on the temporal and spatial dynamic characteristic variables generated in steps S1 and S2, fuse them to obtain the temporal and spatial dynamic characteristic variables of the satellite data, denoted as Latent. TSFN ; S4. Design a Transformer-based neural network, denoted as TSAD. Predict TSAD Predict Based on the spatiotemporal dynamic characteristic variables generated in step (3), predict the values ​​of telemetry parameters after P frames; S5. Use the telemetry data generated by the satellite under normal operating conditions to perform end-to-end training on the neural network model composed of steps S1-S4 to obtain the trained model, denoted as TSAD. S6. Input all samples in the training dataset into the model generated in step (5) and calculate the fault score of all samples in the training set; S7. Based on the set of fault scores generated on the training dataset in step (6), calculate the satellite fault detection threshold using the POT method; S8. Input the test set data into the TSAD model, go through steps S1-S7 in sequence, obtain the detection results of all test set samples, calculate the performance indicators P, R, and F1 of the algorithm on the test set, and repeat steps S1 to S8 to optimize the method performance. S9. Use the satellite data sample to be detected as the model input, generate the final prediction residual from steps (1)-(4), and obtain the final detection result based on the threshold generated in step S7.

2. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, The method for extracting the temporal dynamic features of satellite data in step S1 specifically includes the following steps: S11. Process the raw telemetry data using a sliding window to obtain a single input sample of M consecutive K frames, as shown in the following formula: In the formula: S input Represents a single input sample; X i ,1≤i≤K represents the M-dimensional data of the i-th frame; S12. Design a bidirectional LSTM-based autoencoder neural network TSAD TFN It consists of two parts: an encoder network and a decoder network, denoted as TFN respectively. Encoder and TFN decoder ; S13. Encoder Network TFN Encoder From the output S of step (1.1) input The time feature encoding is extracted using the following formula: Latent TFN =TFN encoder (S input ) In the formula: S input Represents a single input sample; Latent TFN Represents time feature encoding; S14. The decoder network decodes the temporal feature code extracted in step (1.3) using the following formula: O TFN =TFN decoder (Latent TFN ) In the formula: O TFN This represents the data output by the decoder network; S15. Calculate the autoencoder neural network TSAD TFN Input S input With output O TFN The reconstruction error is used as part of the loss function in step (5) to complete the model training, as shown in the following formula:

3. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, The method for extracting spatial dynamic features of satellite data in step S2 specifically includes the following steps: S21. Design a convolution-based autoencoder network TSAD SFN It consists of two parts: an encoder network and a decoder network, denoted as SFN respectively. Encoder and SFN decoder ; S22. Encoder Network SFN Encoder From the output S of step S11 input The spatial feature encoding is extracted from it, and the formula is as follows: Latent SFN =SFN encoder (S input ) In the formula: Latent SFN Represents spatial feature encoding; S23. The decoder network decodes the spatial feature encoding extracted in step S23, using the following formula: O SFN =SFN decoder (Latent SFN ) In the formula: O SFN This represents the data output by the decoder network; S24. Calculate the autoencoder network TSAD SFN Input S input With output O SFN The reconstruction error is used as part of the loss function in step S5 to complete the model training, as shown in the following formula:

4. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, Step S3 fuses the spatiotemporal dynamic features of the satellite data, as shown in the following formula: Latent TSFN =[Latent TFN ,Latent SFN ] In the formula: Latent TSFN This indicates the spatiotemporal dynamic characteristics of the extracted satellite data samples.

5. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, Step S4, which predicts the value of the satellite data after the P-frame, specifically involves: S41. Design a Transformer-based satellite data prediction network TSAD Predict ; S42. The spatiotemporal dynamic features obtained in step S3 are Latent TSFN TSAD (Tracking Satellite Data Prediction Network) Predict The predicted value is obtained using the following formula: O Predict =TSAD Predict (Latent TSFN ) In the formula: O Predict This represents the predicted data after P frames. S43. Calculate the predicted network output O Predict The distance from the actual value after P frames is calculated using the following formula: In the formula: Y P This represents the actual data after the P-frame.

6. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, The method for training the neural network model TSAD in step S5 specifically includes the following steps: S51. The neural network TSAD consists of three subnetworks, namely TSAD TFN TSAD SFN and TSAD Predict ; S52. Select M analog telemetry parameters based on the satellite equipment to be fault detected; S53. Select L consecutive frames of data for the M telemetry parameters under normal satellite operating conditions as the training set, denoted as X. train The formula is as follows: In the formula: L represents the number of frames in the training set data, X i ,1≤i≤L represents the M-dimensional data of the i-th frame; S54. Use the "spectral residual" method to remove training set X train The wild value in the middle, to obtain the new X train ; S55. Calculate the training set X train mean μ train and standard deviation δ train , for X train All samples in the dataset are normalized using the following formula: S56. Set the sliding window width to K and the sliding step size to S, in the training set X train The new training set X is obtained by continuously sliding up. train ; S57. Repeat steps (5.3) to (5.6), outputting the training set X. train Corresponding validation set X valid , verification set X valid The number of samples is X train 20% to 30% of the sample size; S58. Based on the neural network TSAD obtained in steps S15, S24, and S43 TFN TSAD SFN and TSAD Predict The loss of the TSAD model is calculated using the following formula: Loss TSAD =λ1Loss TFN +λ2Loss SFN +λ3Loss Predict In the formula: Loss TSAD Let λ1, λ2, and λ3 represent the loss of the neural network TSAD, and λ3 be hyperparameters representing the loss of the neural network TSAD. TFN TSAD SFN and TSAD Predict In total loss TSAD The weighting percentage of the content; S59. Minimize the loss in step S56. TSAD With the goal of training the model TSAD using the backpropagation algorithm.

7. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, The method for calculating the fault scores of all samples in the training set in step S6 specifically includes the following steps: S61. To predict the network output O Predict The L2 distance between the actual parameter value after P frames is used as the fault score for a single sample. S62. Traverse all data samples in the training set and calculate the fault score set of the model on the training set.

8. The satellite fault detection method based on spatiotemporal dynamic feature fusion according to claim 1, characterized in that, The calculation formula for step S8 is as follows: In the formula: TP represents the number of samples correctly classified as faults, FP represents the number of samples incorrectly classified as faults, TN represents the number of samples correctly classified as normal, and FN represents the number of samples incorrectly classified as normal.

Images