Satellite attitude control system fault diagnosis method based on double confrontation mechanism
Through a fault diagnosis method based on the dual adversarial mechanism, a dual adversarial network is built using sample data of in-orbit satellites and satellite semi-physical air float platforms, which solves the problem of scarce failure data of satellite attitude control system, and achieves efficient and accurate fault diagnosis and widely applicable automated diagnosis.
Patent Information
- Application Number
- CN202510500823.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-21
- Publication Date
- 2025-08-01
AI Technical Summary
The in-orbit fault data of satellite attitude control systems is scarce, which makes it difficult to diagnose faults. The existing methods rely on a large amount of high-quality fault data to train models, and the data scarcity problem has not been effectively solved.
The fault diagnosis method based on the dual adversarial mechanism is adopted, and the sample data of the semi-physical air float platform of the in-orbit satellite and satellite are obtained, and the dual adversarial network is built. The variable generation module, dynamic fusion generator, feature extraction module, domain discriminator, generation discriminator and state classifier are used for training to generate target domain reconstruction data, and feature extraction and state classification are optimized.
It realizes efficient and accurate fault diagnosis of the in-orbit satellite attitude control system, expands the scope of application of fault data, improves the accuracy and efficiency of diagnosis, and is suitable for automated diagnosis of various fault states.
Smart Images
Figure CN120408146A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of fault diagnosis of satellite attitude control systems, and in particular to a fault diagnosis method for satellite attitude control systems based on a dual adversarial mechanism. Background Art
[0002] The satellite attitude control system (ACS) is one of the core subsystems of a satellite, which is used to determine and adjust the pointing (attitude) of the satellite in space to meet mission requirements (such as earth observation, communication antenna orientation, solar panel alignment with the sun, etc.). Its core task is to control the rotation of the satellite around three axes (pitch, yaw, roll) to ensure attitude stability or maneuver according to commands.
[0003] The satellite attitude control system is vulnerable to failures of electronic devices or mechanical components due to radiation, drastic temperature changes, and micrometeorite impacts in space. When a failure occurs in the satellite attitude control system, it will affect the attitude accuracy of the satellite at least, and in the worst case, it will even directly cause the satellite to lose its attitude control ability. Therefore, it is very important to conduct fault diagnosis on the satellite attitude control system to ensure its reliable operation in orbit.
[0004] With the development of artificial intelligence technology, methods using various artificial intelligence networks for fault diagnosis have gradually become the mainstream. However, in traditional fault diagnosis methods, a large number of fault samples are often relied on to train the fault diagnosis model. However, in the practical application of high-tech equipment such as satellite attitude control systems, the complexity of the satellite system and the high cost of testing make it extremely difficult to obtain a large number of high-quality fault data. Therefore, the fault data is very limited, and the problem of data scarcity makes the fault diagnosis of satellite attitude control systems more difficult. Summary of the Invention
[0005] In view of the above problems and technical requirements, this application proposes a fault diagnosis method for satellite attitude control systems based on a dual adversarial mechanism. The technical solution of this application is as follows:
[0006] A fault diagnosis method for satellite attitude control systems based on a dual adversarial mechanism, the fault diagnosis method for satellite attitude control systems includes:
[0007] Obtain target domain sample data of the satellite attitude control system of an on-orbit satellite under different operating states during the historical operation process, and source domain sample data of the satellite attitude control system of a satellite semi-physical air-bearing table under different operating states during the simulated operation process. Each group of sample data includes satellite attitude data and attitude control data. The operating states of the satellite attitude control system include normal states and multiple different fault states;
[0008] Build a dual adversarial network, which includes a variable generation module, a dynamic fusion generator, a feature extraction module, a domain discriminator, a generation discriminator, a domain classifier, and a state classifier;
[0009] Train the dual adversarial network using the target domain sample data and source domain sample data in different operating states: The variable generation module extracts the domain-invariant features and domain-specific features of each group of input sample data, and uses the domain discriminator to distinguish the domain-invariant features of the sample data in different domains to perform adversarial training on the variable generation module, and uses the domain classifier to perform domain classification on the domain-specific features of the sample data in different domains; The dynamic fusion generator generates target domain reconstruction data based on the domain-invariant features and domain-specific features of each group of sample data, and uses the generation discriminator to distinguish the target domain reconstruction data and the target domain sample data to perform adversarial training on the dynamic fusion generator; The feature extraction module extracts features from the input target domain data and uses the state classifier for state classification, and the target domain data includes target domain sample data and target domain reconstruction data;
[0010] Use the feature extraction module and the state classifier in the trained dual adversarial network to perform fault diagnosis on the satellite attitude control system of the on-orbit satellite.
[0011] A further technical solution thereof is that the dynamic fusion generator generates target domain reconstruction data with the same operating state as the current target domain sample data according to the domain-invariant features generated from any group of source domain sample data, the domain-specific features generated from any group of target domain sample data, and the added random noise.
[0012] A further technical solution thereof is that training the dual adversarial network using the target domain sample data and source domain sample data in different operating states includes:
[0013] Weight the self-supervised alignment loss L1 between the source domain and the target domain, the domain discrimination loss L2 of the domain discriminator, the reconstruction discrimination loss L3 of the generation discriminator, the domain classification loss L4 of the domain classifier, and the state classification loss L5 of the state classifier to obtain the total loss function Loss = α1 * L1 + α2 * L2 + α3 * L3 + α4 * L4 + α5 * L5, and adjust the network parameters of the dual adversarial network according to the total loss function Loss for network training; where α1, α2, α3, α4, and α5 are all weighting coefficients and α1 + α2 + α3 + α4 + α5 = 1.
[0014] A further technical solution thereof is that the variable generation module includes a VAE encoder and a latent space decomposition module. The VAE encoder extracts the hidden variables of each group of input sample data, and the latent space decomposition module decomposes the hidden variables into domain-invariant features and domain-specific features.
[0015] Its further technical solution is that the self-supervised alignment loss L1 between the source domain and the target domain = β1 * L rot + β2 * L constract + β3 * L cycle , where:
[0016] The rotation prediction loss L rot is the prediction loss of the rotation-invariant features in the sample data;
[0017] The cross-domain contrast loss L constract is the consistency loss between the domain-invariant features generated by the source domain sample data and the target domain sample data with the same satellite attitude data;
[0018] The cycle consistency loss L cycle is the consistency loss between the domain-invariant features of the target domain reconstruction data extracted by using the variable generation module and the domain-invariant features of the source domain sample data for generating the target domain reconstruction data;
[0019] β1, β2, and β3 are all weighting coefficients and β1 + β2 + β3 = 1.
[0020] Its further technical solution is that the satellite attitude data of each group of sample data includes the satellite rotation angle, and calculating the rotation prediction loss L rot includes:
[0021] Using the self-supervised head to generate a satellite rotation angle prediction value according to the hidden variable of each group of sample data, and calculating the angle loss with the satellite rotation angle in the sample data, and comprehensively obtaining the rotation prediction loss L rot .
[0022] Its further technical solution is that calculating the cross-domain contrast loss L constract includes:
[0023] Calculating the cosine similarity between the domain-invariant features generated by the source domain sample data and the target domain sample data in any sample data combination with the same included satellite attitude data, and comprehensively obtaining the cross-domain contrast loss L constract .
[0024] Its further technical solution is that the dynamic fusion generator generates a group of target domain reconstruction data including:
[0025] Optionally select a set of source domain sample data and a set of target domain sample data, perform linear interpolation on the domain-specific features generated from the source domain sample data and the domain-specific features generated from the target domain sample data to obtain the smoothed domain-specific features of the target domain sample data; then use the domain-invariant features generated from the source domain sample data, the smoothed domain-specific features of the target domain sample data, and the added random noise to generate the target domain reconstruction data.
[0026] A further technical solution thereof is that the fault states of the satellite attitude control system include star sensor faults, gyroscope faults, flywheel faults, actuator constant deviation faults, actuator jamming faults, and control algorithm failure faults.
[0027] A further technical solution thereof is that obtaining the sample data of the satellite attitude control system in each operating state includes:
[0028] Collect the time series of multivariate operating data in the current operating state through the multivariate sensors of the satellite attitude control system;
[0029] After normalizing the time series of multivariate operating data, use a sliding window to sequentially slide and intercept the time series of multivariate operating data, and construct a set of sample data in the current operating state from the time series of multivariate operating data in each sliding window.
[0030] The beneficial technical effects of this application are:
[0031] This application discloses a satellite attitude control system fault diagnosis method based on a dual adversarial mechanism. Based on the dual adversarial network built in this application, the source domain sample data of the satellite attitude control system of the satellite semi-physical air-bearing table in different operating states during the simulation operation can be extended to generate target domain reconstruction data. Thus, combining the target domain reconstruction data and the target domain sample data to jointly train the feature extraction module and the state classifier in the dual adversarial network. The dual adversarial network built in this application uses a dual adversarial mechanism for model training, which can not only ensure the cross-platform consistency between the source domain and the target domain, but also optimize the authenticity of the target domain reconstruction data, thereby making up for the defect of limited fault data of the satellite attitude control system of the on-orbit satellite. So that the feature extraction module and the state classifier trained by the dual adversarial network can be used for the automatic fault diagnosis of the satellite attitude control system of the online satellite, with high diagnostic accuracy and efficiency, and a wide range of applications.
[0032] When constructing the loss function to train the dual adversarial network, in addition to designing the losses of each classifier and discriminator, when calculating the self-supervised alignment loss between the source domain and the target domain, not only the cross-domain contrast loss and the cycle consistency loss are added to optimize the cross-platform consistency, but also the rotation prediction loss is added to optimize the accurate extraction of the core feature of the satellite rotation angle in the satellite attitude angular velocity by the network, so as to further optimize the accuracy and robustness of the feature extraction module and the state classifier in the trained dual adversarial network. Brief Description of the Drawings
[0033] Figure 1 is a schematic flowchart of a satellite attitude control system fault diagnosis method according to an embodiment of the present application.
[0034] Figure 2 is a schematic diagram of network training using the dual adversarial network built by the present application. Detailed Embodiments
[0035] The following further describes the detailed embodiments of the present application with reference to the accompanying drawings.
[0036] The present application discloses a satellite attitude control system fault diagnosis method based on a dual adversarial mechanism. Please refer to Figure 1 the flowchart shown, and the satellite attitude control system fault diagnosis method includes the following:
[0037] Step 1, obtain target domain sample data of the satellite attitude control system of the on-orbit satellite in different operating states during the historical operation process, and source domain sample data of the satellite attitude control system of the satellite semi-physical air-bearing table in different operating states during the simulated operation process.
[0038] For each group of sample data in the target domain sample data and the source domain sample data, the data content of a group of sample data includes satellite attitude data and attitude control data. The satellite attitude data includes satellite attitude angles, satellite attitude angular velocities, flywheel angular accelerations, etc., and the attitude control data includes control torques, etc. The operating states of the satellite attitude control system include normal states and various different fault states. The sample data in a resulting operating state has a label corresponding to the operating state. The fault states targeted by the present application include star sensor faults, gyroscope faults, flywheel faults, actuator constant deviation faults, actuator jamming faults, and control algorithm failure faults.
[0039] The methods for obtaining target domain sample data and source domain sample data are similar. The method for obtaining any set of sample data includes: collecting the time series of multivariate operation data in the current operating state through the multivariate sensors of the satellite attitude control system, first performing normalization processing on the time series of multivariate operation data to improve the training stability of the subsequent model. Then, use a sliding window to sequentially slide and intercept the time series of multivariate operation data, and construct a set of sample data in the current operating state from the time series of multivariate operation data within each sliding window. The satellite semi-physical air-bearing platform can be used to simulate various operating states to obtain a large amount of source domain sample data, but the number of target domain sample data of the on-orbit satellite is relatively limited.
[0040] Step 2, build a dual adversarial network. Please refer to Figure 1 the network architecture shown. The dual adversarial network built in this application includes a variable generation module, a dynamic fusion generator, a feature extraction module, a domain discriminator, a generation discriminator, a domain classifier, and a state classifier. These functional modules are respectively implemented based on existing deep neural network architectures, such as being built based on the LSTM network.
[0041] Step 3, use the target domain sample data and source domain sample data in different operating states to train the dual adversarial network built in this application. Please combine Figure 2 with the network architecture diagram. The training process is as follows:
[0042] Each set of sample data in the target domain sample data and the source domain sample data will first be input into the variable generation module, and the variable generation module extracts the domain-invariant features and domain-specific features of the input each set of sample data. In one embodiment, as Figure 2 shown, the variable generation module includes a VAE encoder and a latent space decomposition module. The VAE encoder extracts the hidden variables of the input each set of sample data, and the latent space decomposition module decomposes the hidden variables into domain-invariant features and domain-specific features. As Figure 2 shown, for the convenience of distinction, the domain-invariant features generated by the variable generation module for the input source domain sample data Data_A will be denoted as Z SA , the domain-specific features will be denoted as Z DA , and the domain-invariant features generated by the variable generation module for the input target domain sample data Data_B will be denoted as Z SB , and the domain-specific features will be denoted as Z DB .
[0043] The domain-invariant features of each group of sample data generated by the variable generation module are input into the domain discriminator. The domain discriminator is used to distinguish whether the domain-invariant features belong to the target domain or the source domain, and the output is the probability of belonging to the target domain (or source domain). Thus, the domain discriminator is used to distinguish the domain-invariant features of sample data in different domains to perform adversarial training on the variable generation module, thereby optimizing the representation ability of the extracted domain-invariant features in the target domain and the source domain and improving the extraction ability of the variable generation module for domain-invariant features.
[0044] The domain-specific features of each group of sample data generated by the variable generation module are input into the domain classifier. The domain classifier is used to classify the input sample data according to the domain-specific features, and the output is the probability distribution of belonging to the target domain and the source domain.
[0045] After the variable generation module generates the domain-invariant features and domain-specific features of each group of sample data respectively, the dynamic fusion generator generates the target domain reconstruction data according to the domain-invariant features and domain-specific features of each group of sample data. Specifically: any two combinations of the target domain sample data and the source domain sample data form a sample data group, so as to construct multiple sample data groups. As Figure 2 shown, the domain-invariant feature Z SA generated from the source domain sample data in the same sample data group and the domain-specific feature Z DB generated from the target domain sample data are input into the dynamic fusion generator. The dynamic fusion generator generates the target domain reconstruction data with the same operating state as the current target domain sample data according to the domain-invariant feature Z SA generated from the source domain sample data in the same sample data group, the domain-specific feature Z DB generated from the target domain sample data, and the added random noise. The addition of random noise can improve data stability.
[0046] Furthermore, in order to achieve a smooth transition between the source domain and the target domain and thus reduce the impact brought by the platform difference between the on-orbit satellite and the satellite semi-physical air-bearing platform, the domain-specific feature Z DA generated from the source domain sample data in the same sample data combination and the domain-specific feature Z DB generated from the target domain sample data are linearly interpolated to obtain the smoothly transitioned domain-specific feature Z' DB of the target domain sample data = λ * Z DA +(1 - λ) * Z DB , where λ is an interpolation coefficient in the range of 0 to 1. Then, the domain-invariant feature Z SA generated from the source domain sample data in the sample data combination, the smoothly transitioned domain-specific feature Z' DB of the target domain sample data, and the added random noise are used to generate the target domain reconstruction data.
[0047] The generated target domain reconstruction data and the initial target domain sample data are merged to form target domain data. On the one hand, the obtained target domain data is input into the generation discriminator, and the generation discriminator discriminates each group of input target domain data to determine whether it belongs to the target domain reconstruction data or the target domain sample data, so as to perform adversarial training on the dynamic fusion generator to optimize the dynamic fusion generator to generate more realistic target domain reconstruction data. On the other hand, the obtained target domain data is input into the feature extraction module, and the feature extraction module extracts features based on the input target domain data and performs state classification using the state classifier.
[0048] The target domain sample data and the source domain sample data are input into the double adversarial network for post-processing according to the above process, and the total loss function Loss is calculated. The total loss function Loss of this application includes the following items:
[0049] (1) The self-supervised alignment loss L1 between the source domain and the target domain, which represents the content alignment between the feature distributions of the source domain sample data and the target domain sample data. The self-supervised alignment loss L1 between the source domain and the target domain = β1 * L rot + β2 * L constract + β3 * L cycle , where β1, β2, and β3 are all weighting coefficients and β1 + β2 + β3 = 1. The other losses include:
[0050] (a) The rotation prediction loss L rot is the prediction loss for the rotation-invariant features in each group of sample data.
[0051] As described above, each group of sample data includes satellite attitude data, and the satellite attitude data includes satellite attitude angles. The satellite rotation angle in the satellite attitude angles is the core data in the actual satellite operation process. Therefore, the self-supervised head is used to generate the satellite rotation angle prediction value according to the hidden variables of each group of sample data, and the angle loss between the satellite rotation angle prediction value and the satellite rotation angle in the sample data is calculated. The rotation prediction loss L is obtained by integrating the angle losses of all sample data. rot . The rotation prediction loss L <> rot helps to optimize the accurate extraction of the core features in the satellite attitude data, that is, the satellite rotation angle, by the network.
[0052] (b) The cross-domain contrast loss L constractis the consistency loss between the domain-invariant features generated from the source-domain sample data and the target-domain sample data with the same satellite attitude data. Specifically: Calculate the cosine similarity between the domain-invariant features generated from the source-domain sample data and the target-domain sample data in any sample data combination containing the same satellite attitude data, and obtain the cross-domain contrast loss L by integrating the average value of the cosine similarities of each sample data combination with the same satellite attitude data. constract . The cross-domain contrast loss L constract helps to improve the cross-platform consistency between the domain-invariant features of the source-domain sample data and the domain-invariant features of the target-domain sample data.
[0053] (c) The cycle consistency loss L cycle is the consistency loss between the domain-invariant features of the target-domain reconstructed data extracted by the variable generation module and the domain-invariant features of the source-domain sample data that generates the target-domain reconstructed data. The cycle consistency loss L cycle also helps to improve the cross-platform consistency of the data.
[0054] (2) The domain discrimination loss L2 of the domain discriminator, which represents the loss between the domain discrimination result of the domain discriminator on the domain-invariant features of each group of sample data and the actual domain to which the sample data belongs.
[0055] (3) The reconstruction discrimination loss L3 of the generation discriminator, which represents the loss between the discrimination result of the generation discriminator on whether a group of target-domain data belongs to the target-domain sample data / target-domain reconstructed data and the actual target-domain sample data / target-domain reconstructed data to which the target-domain data belongs.
[0056] (4) The domain classification loss L4 of the domain classifier, which represents the loss between the probability distribution of the domain to which the domain-specific features of a group of sample data belong and the actual domain label of the sample data.
[0057] (5) The state classification loss L5 of the state classifier, which represents the loss between the classification result of the state classifier on the operating state of a group of target-domain data and the label of the actual operating state of the target-domain data.
[0058] The domain discrimination loss L2, reconstruction discrimination loss L3, domain classification loss L4, and state classification loss L5 can be calculated based on existing formulas, such as the cross-entropy loss. The total loss function Loss = α1*L1+α2*L2+α3*L3+α4*L4+α5*L5 can then be calculated, where α1, α2, α3, α4, and α5 are weighting coefficients and α1+α2+α3+α4+α5=1. The network parameters of the dual adversarial network are then adjusted according to the total loss function Loss, and the network is trained until convergence. During this training process, the dual discriminant structure consisting of the domain discriminator and the generative discriminator forms a dual adversarial mechanism, enabling the source domain sample data of the satellite semi-physical air-floating platform and the target domain sample data of the on-orbit satellite to be mapped to the same domain, achieving auxiliary training effects for the source domain sample data of the satellite semi-physical air-floating platform.
[0059] Step 4. After the network training is completed, the feature extraction module and state classifier in the trained dual adversarial network can be used to perform fault diagnosis on the satellite attitude control system of the on-orbit satellite: extract the multivariable operation data time series of the satellite attitude control system of the on-orbit satellite during the actual operation process and input it into the feature extraction module for feature extraction, and then use the state classifier to classify and output the classification result of the operation state, thereby realizing the fault diagnosis of the satellite attitude control system of the on-orbit satellite.
[0060] The above description is only a preferred embodiment of the present application, and the present application is not limited to the above embodiments. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present application should be considered to be included in the scope of protection of the present application.
Claims
1. A fault diagnosis method for a satellite attitude control system based on a dual adversarial mechanism, characterized in that, The satellite attitude control system fault diagnosis method includes: Obtaining target domain sample data of the satellite attitude control system of an on-orbit satellite under different operating states during the historical operation process, and source domain sample data of the satellite attitude control system of a satellite semi-physical air-bearing table under different operating states during the simulated operation process. Each set of sample data includes satellite attitude data and attitude control data. The operating states of the satellite attitude control system include a normal state and multiple different fault states; Building a dual adversarial network, where the dual adversarial network includes a variable generation module, a dynamic fusion generator, a feature extraction module, a domain discriminator, a generation discriminator, a domain classifier, and a state classifier; Training the dual adversarial network using the target domain sample data and source domain sample data under different operating states: The variable generation module extracts the domain-invariant features and domain-specific features of each input set of sample data, and uses the domain discriminator to distinguish the domain-invariant features of sample data in different domains to perform adversarial training on the variable generation module, and uses the domain classifier to perform domain classification on the domain-specific features of sample data in different domains; The dynamic fusion generator generates target domain reconstruction data based on the domain-invariant features and domain-specific features of each set of sample data, and uses the generation discriminator to distinguish the target domain reconstruction data and the target domain sample data to perform adversarial training on the dynamic fusion generator; The feature extraction module extracts features from the input target domain data and uses the state classifier to perform state classification. The target domain data includes target domain sample data and target domain reconstruction data; Using the feature extraction module and state classifier in the trained dual adversarial network to perform fault diagnosis on the satellite attitude control system of the on-orbit satellite.
2. The satellite attitude control system fault diagnosis method according to claim 1, wherein The dynamic fusion generator generates target domain reconstruction data with the same operating state as the current target domain sample data based on the domain-invariant features generated from any set of source domain sample data, the domain-specific features generated from any set of target domain sample data, and added random noise.
3. The satellite attitude control system fault diagnosis method according to claim 2, wherein Training the dual adversarial network using the target domain sample data and source domain sample data under different operating states includes: Weighting the self-supervised alignment loss L1 between the source domain and the target domain, the domain discrimination loss L2 of the domain discriminator, the reconstruction discrimination loss L3 of the generation discriminator, the domain classification loss L4 of the domain classifier, and the state classification loss L5 of the state classifier to obtain the total loss function Loss = α1*L1 + α2*L2 + α3*L3 + α4*L4 + α5*L5, and adjusting the network parameters of the dual adversarial network according to the total loss function Loss for network training; where α1, α2, α3, α4, and α5 are all weighting coefficients and α1 + α2 + α3 + α4 + α5 = 1.
4. The satellite attitude control system fault diagnosis method according to claim 3, wherein, The variable generation module includes a VAE encoder and a latent space decomposition module. The VAE encoder extracts the hidden variables of each input set of sample data, and the latent space decomposition module decomposes the hidden variables into domain-invariant features and domain-specific features.
5. The satellite attitude control system fault diagnosis method according to claim 4, wherein The self-supervised alignment loss L1 between the source domain and the target domain = β1 * L rot + β2 * L constract + β3 * L cycle , where: Rotation prediction loss L rot is the prediction loss for rotation-invariant features in the sample data; Cross-domain contrastive loss L constract is the consistency loss between the domain-invariant features generated by the source-domain sample data and the target-domain sample data with the same satellite attitude data; Cyclic consistency loss L cycle is the consistency loss between the domain-invariant features of the target-domain reconstructed data extracted by the variable generation module and the domain-invariant features of the source-domain sample data for generating the target-domain reconstructed data; β1, β2, and β3 are all weighting coefficients and β1 + β2 + β3 = 1.
6. The satellite attitude control system fault diagnosis method according to claim 5, wherein The satellite attitude data of each group of sample data includes the satellite rotation angle, and the rotation prediction loss L is calculated rot including: Using a self-supervised head to generate a predicted value of the satellite rotation angle based on the hidden variable of each set of sample data, calculating the angular loss with the satellite rotation angle in the sample data, and obtaining the rotation prediction loss L by integrating the angular losses of all sample data rot .
7. The satellite attitude control system fault diagnosis method according to claim 6, wherein, Calculate the cross-domain contrastive loss L constract including: Calculate the cosine similarity between the domain-invariant features generated by the source domain sample data and the target domain sample data in any sample data combination with the same included satellite attitude data, and obtain the cross-domain contrast loss L by integrating the average value of the cosine similarities of each sample data combination with the same satellite attitude data constract .
8. The satellite attitude control system fault diagnosis method according to claim 2, characterized in that The dynamic fusion generator generating a set of target domain reconstruction data includes: Optionally select a set of source domain sample data and a set of target domain sample data, perform linear interpolation on the domain-specific features generated from the source domain sample data and the domain-specific features generated from the target domain sample data to obtain the smoothed domain-specific features of the target domain sample data; then use the domain-invariant features generated from the source domain sample data, the smoothed domain-specific features of the target domain sample data, and the added random noise to generate the target domain reconstruction data.
9. The satellite attitude control system fault diagnosis method according to claim 1, characterized in that The fault states of the satellite attitude control system include star sensor faults, gyroscope faults, flywheel faults, actuator constant deviation faults, actuator jamming faults, and control algorithm failure faults.
10. The satellite attitude control system fault diagnosis method according to claim 1, wherein Obtaining the sample data of the satellite attitude control system in each operating state includes: Collecting the time series of multivariate operating data in the current operating state through the multivariate sensors of the satellite attitude control system; After normalizing the time series of multivariate operating data, use a sliding window to sequentially slide and intercept the time series of multivariate operating data, and construct a set of sample data in the current operating state from the time series of multivariate operating data in each sliding window.