Satellite attitude control system fault detection method, equipment and medium

Abstract

The invention discloses a satellite attitude control system fault detection method and device and a medium, and the method comprises the steps: introducing a diffusion model to carry out the preprocessing of an original data set before a fault diagnosis task is executed through employing a neural network, carrying out the focusing training of an efficient diffusion model, generating scarce minority class samples through employing the model, and carrying out the fault diagnosis of a satellite attitude control system. The problem of category imbalance generally existing in satellite attitude control system data is effectively solved, and the strong data processing capacity of the diffusion model is also utilized. Based on the balanced training set, the adaptive one-dimensional convolutional neural network is further combined, complex time sequence data is efficiently processed, key fault information is mined, and the accuracy and reliability of fault diagnosis of the satellite attitude control system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite fault detection, and in particular to a method, device and medium for detecting faults in a satellite attitude control system. Background Art

[0002] Since the launch of the world's first satellite, mankind has ushered in a new era of space exploration. The application of satellites is no longer limited to the military and political fields, but has widely penetrated into civilian fields such as navigation, detection, communication, and meteorology, bringing huge economic benefits and convenience to daily life for mankind. Today, artificial satellites play a vital role in the development of many industries. There are many types of satellites, and their structures are becoming increasingly complex. They are usually composed of multiple complex subsystems such as attitude control, communication, and power supply. As a large-scale, complex and expensive system, the reliability of satellites is the cornerstone of their normal operation. However, due to the high complexity and nonlinear characteristics of satellite systems, coupled with the limitations of process level and capital cost, as well as the harsh on-orbit working environment, such as frequent and large changes in high and low temperatures and strong cosmic electromagnetic radiation, various subsystems and components are prone to failure during on-orbit operation.

[0003] The consequences of satellite failure may be very serious, ranging from huge economic losses to catastrophic consequences. The satellite attitude control system has a high probability of failure during operation. Minor failures may weaken the system accuracy and performance, while serious failures may cause the satellite mission to fail, which will not only cause huge manpower and economic losses, but may also become "space junk". Therefore, fault diagnosis of the satellite's attitude control system is of great significance to improving the stability and healthy operation capability of the satellite.

[0004] During the normal on-orbit operation of the satellite, the amount of data collected by the sensors and actuators on the working status of the reaction system is extremely large, just for the attitude control system. Even if these data are screened and only valid information is retained, the amount of data is still astonishing after a long period of accumulation. As the data transmitted to the ground station by the satellite during its on-orbit operation becomes more and more stable, the proportion of fault data on the satellite attitude in the massive telemetry data gradually decreases, and the number of samples becomes limited, resulting in data imbalance.

[0005] Therefore, it is necessary to effectively diagnose the faults of satellite attitude control systems. At present, although many data-driven fault diagnosis technologies have been widely used in satellite attitude control systems, such as algorithms based on convolutional neural networks, models based on long short-term memory networks, and methods using autoencoders, these technologies have extremely strict requirements on data, mainly including: traditional methods may not meet real-time requirements in terms of diagnostic accuracy and speed, especially for fast and accurate diagnosis in complex environments. Although some deep learning-based methods, such as deep belief networks (DBNs) and stacked denoising autoencoders, have achieved good results, they have problems with complex network structures and large computational complexity, which may limit their application in resource-constrained satellite environments. Existing methods have limited ability to handle modeling uncertainty and nonlinear systems. Summary of the invention

[0006] The technical problem to be solved by the present invention is that the efficiency and accuracy of existing fault diagnosis are low. The purpose is to provide a satellite attitude control system fault detection method, device and medium. By introducing a diffusion model to pre-process the original data set before using a neural network to perform a fault diagnosis task, not only the category imbalance problem commonly existing in satellite attitude control system data is effectively solved, but also the representativeness and generalization of the data set are enhanced through the powerful data processing capability of the diffusion model. Combined with an adaptive one-dimensional convolutional neural network, fault information is identified, thereby improving the accuracy and reliability of satellite attitude control system fault diagnosis.

[0007] The present invention is achieved through the following technical solutions:

[0008] A first aspect of the present invention provides a satellite attitude control system fault detection method, comprising the following specific steps:

[0009] S1: Acquire key data in the satellite attitude control system, wherein the key data includes real-time operating parameters of the satellite attitude control components;

[0010] S2: Normalize the key data obtained in the satellite attitude control system and divide the normalized data into training set and test set;

[0011] S3: Construct a diffusion model, use the fault data in the training set to train the diffusion model, and use the trained diffusion model to balance the normal data and fault data in the training set;

[0012] S4: construct an adaptive one-dimensional neural network classifier, and use the balanced training set to train the adaptive one-dimensional neural network classifier;

[0013] S5: The trained adaptive one-dimensional neural network classifier is evaluated using the normalized test set.

[0014] Furthermore, the normalization of the key data acquired in the satellite attitude control system specifically includes:

[0015]

[0016] Among them, norm represents the normalized result, x i represents the i-th sample, max(x) represents the maximum value in the sample, and min(x) represents the minimum value.

[0017] Furthermore, the step S3 specifically includes:

[0018] S31: Build and train the diffusion model to generate new fault samples;

[0019] S32: Merge the new fault sample with the training set to obtain a balanced training set.

[0020] Furthermore, the step S31 specifically includes:

[0021] Define a forward diffusion process to gradually inject random noise into the fault data in the training set until the original fault data is transformed from the original distribution to a Gaussian distribution;

[0022] The reverse diffusion process is defined to gradually denoise the data with standard Gaussian distribution and generate new data.

[0023] Furthermore, the definition of the forward diffusion process specifically includes:

[0024] Obtain the fault data in the training set as sample data;

[0025] Obtain a noise time table, where the noise time table is a parameter value for adding noise at time t, and is used to control the amount of noise added to the sample data;

[0026] Define the noise addition time step t, and at each time step t, calculate an intermediate variable based on the noise variance;

[0027] The sample data is subjected to noise processing based on the intermediate variables and the noise time table. The noise processing includes n time steps, each time step generates a noise process data, and each time step performs superimposed noise processing on the noise process data generated by the previous time step until a predetermined number of steps is reached, the sample data is transformed from the original distribution to the Gaussian distribution, and the noise processing is terminated.

[0028] Furthermore, the definition of the reverse diffusion process specifically includes:

[0029] Construct the probability distribution of the reverse diffusion process and construct the loss function of the neural network for predicting noise by combining the posterior conditional probability;

[0030] The loss function is optimized using a variational lower bound, a neural network is constructed, and the neural network is trained using data from the forward process;

[0031] Based on the trained neural network, the noise is predicted once at each time step;

[0032] Based on the predicted noise, the mean of the probability distribution of the reverse diffusion process is calculated to obtain the probability distribution of the reverse diffusion process;

[0033] According to the probability distribution, the data of the standard Gaussian distribution is denoised once at each time step until the predetermined number of steps is reached to generate new data.

[0034] Furthermore, the step S4 specifically includes:

[0035] Constructing an adaptive one-dimensional convolutional neural network, wherein the input of the adaptive one-dimensional convolutional neural network is a flywheel output torque signal of a satellite attitude control system, and the output is a fault classification number;

[0036] Set parameters for the adaptive one-dimensional convolutional neural network;

[0037] A loss function is defined based on the cross entropy error, and an adaptive one-dimensional convolutional neural network is optimized based on the loss function;

[0038] The balanced training set is used to train the adaptive one-dimensional convolutional neural network.

[0039] Furthermore, the test set obtained after normalization is used to evaluate the trained adaptive one-dimensional neural network classifier, specifically including:

[0040]

[0041] Among them, F1 represents the weighted average evaluation score of precision and recall, P represents precision, R represents recall, TP represents true positive examples, FP represents false positive examples, and FN represents false negative examples.

[0042] A second aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for detecting faults in a satellite attitude control system is implemented.

[0043] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a satellite attitude control system fault detection method.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] The present invention introduces a diffusion model to pre-process the original data set before using a neural network to perform fault diagnosis tasks. This not only effectively solves the class imbalance problem that is prevalent in satellite attitude control system data, but also enhances the representativeness and generalization of the data set through the powerful data processing capability of the diffusion model. Combined with an adaptive one-dimensional convolutional neural network, the fault information is identified, thereby improving the accuracy and reliability of satellite attitude control system fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0047] Figure 1 A schematic diagram of a method flow in an embodiment of the present invention;

[0048] Figure 2 4 is a structural diagram of a diffusion model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0050] As a possible implementation, Figure 1 As shown, this embodiment provides a satellite attitude control system fault detection method based on a diffusion model. The method first focuses on training an efficient diffusion model and uses this model to generate scarce minority class samples, thereby effectively solving the problem of class imbalance in the data set and achieving the balance of the data set. Subsequently, based on the balanced data set, an adaptive one-dimensional convolutional neural network is further combined to efficiently process complex time series data and mine key fault information to significantly improve its accuracy and efficiency in satellite attitude fault detection tasks. Through this series of steps, the present invention not only optimizes the data processing process, but also enhances the reliability of fault detection.

[0051] The specific solutions of this embodiment include:

[0052] S1: The ground control station collects key data from the satellite attitude control system, including real-time operating parameters of key components such as flywheels and gyroscopes;

[0053] S2: Normalize the satellite attitude control system data collected by S1, and divide the normalized data into a training set and a test set;

[0054] S3: Construct a diffusion model, use the training set to train the diffusion model, and use the trained diffusion model to balance the normal data and fault data in the training set in S2 to balance the training set;

[0055] S4: construct an adaptive one-dimensional neural network classifier, and use the balanced training set to train the adaptive one-dimensional neural network classifier;

[0056] S5: The trained adaptive one-dimensional neural network classifier is evaluated using the normalized test set.

[0057] In some possible implementations, step S2 specifically includes:

[0058] S21: The data normalization formula is:

[0059]

[0060] Among them, norm represents the normalized result, x i represents the i-th sample, max(x) represents the maximum value in the sample, and min(x) represents the minimum value.

[0061] S22: The normalized data can be divided into a training set and a test set in a ratio of 8:2.

[0062] In some possible implementations, step S3 specifically includes:

[0063] S31: Construct a diffusion model. The denoising diffusion probability model (DDPM) is a probability generation model that includes two stages: forward and reverse. It is usually assumed that both the forward and reverse stages are Markov processes. The model structure is as follows: Figure 2 shown.

[0064] In the forward phase of DDPM, also known as the forward diffusion process or the forward noisy process, the diffusion model gradually injects random noise into the key data (i.e., sample data) collected in the satellite attitude control system until the data is completely converted into random noise.

[0065] Assuming the initial distribution x0~q(x), the noise adding operation is divided into T steps to add noise to the sample data. At each moment t in the forward process, Gaussian noise with specific mean and standard deviation is added to the data.

[0066]

[0067] Among them, β t are the artificially set noise parameter values ​​at time t, β1, β2, ..., β T Construct a noise schedule (noiseschedule), and 0<β1<...<β t-1 <β t <...<β T <1, t represents the time step.

[0068] Get x0 to x t The transition probability, that is, from the initial state x0 to the state x at time t t The conditional probability distribution of is:

[0069]

[0070] Where N represents the normal distribution, represents the mean, represents the covariance matrix, I is the identity matrix, α t =1-β t ,a t represents the intermediate variable, Indicates a s Cumulative effect from time 0 to t;

[0071] x t Reparameterization:

[0072]

[0073] From the above formula, we can see that in the whole forward process, for the determined sample data x0 and noise time table β1, β2...β T , can get the data x of any step at one time t At the same time, when the total number of steps T is large enough, the final noise result x T can be regarded as a Gaussian distribution, that is

[0074] In contrast, the reverse phase plays the role of data reconstruction, also known as the reverse generation process or reverse denoising process. In this phase, the model tries to obtain clear data samples from pure noise by gradually removing noise.

[0075] The reverse process usually uses a neural network parameterized by θ to approximate its distribution. Assume that p θ (x t-1 ∣x t ) is the probability distribution of the reverse process and obeys the Gaussian distribution. Considering that from x t Remove the noise estimate x from step tt-1 , whose mean value μ θ (x t ,t) and variance can be regarded as x t And the function of t:

[0076]

[0077] Among them, the variance is set to a constant that does not need to participate in neural network training and is time-dependent. Therefore, during training, only the neural network training mean μ is used θ (x t ,t).

[0078]

[0079] Among them, ε θ is the predicted value of the noise ε.

[0080] In order to determine the neural network training μ θ (x t ,t), we need to use q(x t-1 ∣x t ). Although it is not possible to directly calculate q(x ι-1 ∣x ι ), but can be based on the process value x at time t t and the initial value x0 to calculate the posterior conditional probability q(x ι-1 ∣x ι ,x ° ):

[0081]

[0082] in,

[0083]

[0084] According to p θ (x t-1 ∣x t ),q(x t-1 ∣x t ,x0), and by comparing formula (5) and formula (6), the loss function can be determined.

[0085] Using a neural network to train the mean μ θ (x t , t), there are three options for the amount to be predicted to obtain the training results. One of them is to predict the noise ε, which is converted by formula (4) and then substituted into formula (8) to eliminate x0, and then obtain formula (6) to calculate μ θ (x t ,t). Its loss function is:

[0086]

[0087] Using the variational lower bound to optimize the loss function, we can get:

[0088]

[0089] in, represents the expected value of x0 and ε. Here x0 usually represents the input data, ε represents the prediction noise, ε is a random variable, and it is usually assumed to obey the normal distribution N(0,I), that is, the mean is 0, the covariance matrix is ​​the normal distribution of the identity matrix, ε θ (x t ,t) represents the error of model prediction, and θ represents the model parameters.

[0090] According to the loss function determined by (11), a diffusion model is constructed. The network structure adopts 1D-U-Net network structure. The direct input of the network is x t , after encoding using Embedding technology at time t, the merged residual structure is input into the network, and the output of the network is ε θ ,The number of channels and size are the same as the input.

[0091] S32: Use the fault samples in the training set to train the diffusion model.

[0092] S33: Generate fault samples using the trained diffusion model, and merge the newly generated fault samples with the previous training set.

[0093] In some possible implementations, step S4 specifically includes:

[0094] S41: Construct an adaptive one-dimensional convolutional neural network. Compared with some commonly used fault diagnosis algorithms, the adaptive one-dimensional convolutional neural network has excellent performance in diagnostic accuracy and diagnostic speed. The input of the adaptive one-dimensional convolutional neural network is the flywheel output torque signal of the satellite attitude control system, and the output is the fault classification number. Specifically, the adaptive one-dimensional neural network topology has no restriction on the input signal size. In this embodiment, the adaptive one-dimensional convolutional neural network structure includes 3 convolutional layers and 2 MLP layers. The convolution kernel and pooling factor of the one-dimensional convolutional neural network are set to 9 and 4 respectively, and the pooling factor of the last convolutional layer is set to 4. The loss function is defined based on the cross entropy error, and the adaptive one-dimensional convolutional neural network is optimized based on the loss function, wherein the loss function is:

[0095]

[0096] Among them, CEE is the cross entropy error; p represents the input, is the target output corresponding to p, is the actual output.

[0097] S42: Use the balanced training set to train an adaptive one-dimensional convolutional neural network.

[0098] In some possible implementations, step S5 specifically includes:

[0099] S51: Use the test set to evaluate the fault classification effect of the neural network, and the test indicator is F1-Score.

[0100]

[0101] Among them, F1 represents the weighted average evaluation score of precision and recall, P represents precision, R represents recall, TP represents true positive examples, FP represents false positive examples, and FN represents false negative examples.

[0102] As a possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a satellite attitude control system fault detection method is implemented.

[0103] As a possible implementation, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, a method for detecting faults in a satellite attitude control system is implemented.

[0104] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A satellite attitude control system fault detection method, characterized in that: The specific steps include: S1: Acquire key data in the satellite attitude control system, wherein the key data includes real-time operating parameters of the satellite attitude control components; S2: Normalize the key data obtained in the satellite attitude control system and divide the normalized data into training set and test set; S3: Construct a diffusion model, use the fault data in the training set to train the diffusion model, and use the trained diffusion model to balance the normal data and fault data in the training set; S4: construct an adaptive one-dimensional neural network classifier, and use the balanced training set to train the adaptive one-dimensional neural network classifier; S5: The trained adaptive one-dimensional neural network classifier is evaluated using the normalized test set.

2. The satellite attitude control system fault detection method according to claim 1, characterized in that: The normalization of the key data acquired in the satellite attitude control system specifically includes: Among them, norm represents the normalized result, x i represents the i-th sample, max(x) represents the maximum value in the sample, and min(x) represents the minimum value.

3. The satellite attitude control system fault detection method according to claim 1, characterized in that: The step S3 specifically includes: S31: Build and train the diffusion model to generate new fault samples; S32: Merge the new fault sample with the training set to obtain a balanced training set.

4. The satellite attitude control system fault detection method according to claim 3, characterized in that: The step S31 specifically includes: Define a forward diffusion process to gradually inject random noise into the fault data in the training set until the original fault data is transformed from the original distribution to a Gaussian distribution; Define the reverse diffusion process, gradually denoise the data with standard Gaussian distribution and generate new data.

5. The satellite attitude control system fault detection method according to claim 4, characterized in that: The definition of the forward diffusion process specifically includes: Obtain the fault data in the training set as sample data; Obtain a noise time table, where the noise time table is a parameter value for adding noise at time t, and is used to control the amount of noise added to the sample data; Define the noise addition time step t, and at each time step t, calculate an intermediate variable based on the noise variance; The sample data is subjected to noise processing based on the intermediate variables and the noise time table. The noise processing includes n time steps, each time step generates a noise process data, and each time step performs superimposed noise processing on the noise process data generated by the previous time step until a predetermined number of steps is reached, the sample data is transformed from the original distribution to the Gaussian distribution, and the noise processing is terminated.

6. The satellite attitude control system fault detection method according to claim 5, characterized in that: The definition of the reverse diffusion process specifically includes: Construct the probability distribution of the reverse diffusion process and construct the loss function of the neural network for predicting noise by combining the posterior conditional probability; The loss function is optimized using a variational lower bound, a neural network is constructed, and the neural network is trained using data from the forward process; Based on the trained neural network, the noise is predicted once at each time step; Based on the predicted noise, the mean of the probability distribution of the reverse diffusion process is calculated to obtain the probability distribution of the reverse diffusion process; According to the probability distribution, the data of the standard Gaussian distribution is denoised once at each time step until the predetermined number of steps is reached to generate new data.

7. The satellite attitude control system fault detection method according to claim 1, characterized in that: The step S4 specifically includes: Constructing an adaptive one-dimensional convolutional neural network, wherein the input of the adaptive one-dimensional convolutional neural network is a flywheel output torque signal of a satellite attitude control system, and the output is a fault classification number; Set parameters for the adaptive one-dimensional convolutional neural network; A loss function is defined based on the cross entropy error, and an adaptive one-dimensional convolutional neural network is optimized based on the loss function; The balanced training set is used to train the adaptive one-dimensional convolutional neural network.

8. The satellite attitude control system fault detection method according to claim 1, characterized in that: The test set obtained after normalization is used to evaluate the trained adaptive one-dimensional neural network classifier, specifically including: Among them, F1 represents the weighted average evaluation score of precision and recall, P represents precision, R represents recall, TP represents true positive examples, FP represents false positive examples, and FN represents false negative examples.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the satellite attitude control system fault detection method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the satellite attitude control system fault detection method as described in any one of claims 1 to 8 is implemented.

Images