Satellite attitude control system fault diagnosis method and device based on time-frequency domain feature fusion

Abstract

The present disclosure provides a satellite attitude control system fault diagnosis method and device based on time-frequency domain feature fusion, belonging to the technical field of satellite fault diagnosis. The method can include: obtaining target satellite data and converting the target satellite data into satellite image data using a Gram angle field; performing feature extraction on the satellite image data to obtain a satellite feature image; extracting global time sequence features of the satellite feature image, wherein the global time sequence features refer to image sequence features extracted from the entire time sequence of the satellite and capable of representing the characteristics of the entire sequence; inputting the global time sequence features into a dense connection model to obtain a fusion time sequence feature map corresponding to the satellite feature image; and obtaining a target fault type corresponding to the target satellite data according to the fusion time sequence feature map. This technical solution can improve the efficiency and accuracy of satellite fault diagnosis.

Description

Technical Field

[0001] This disclosure relates to the field of satellite fault diagnosis technology, and in particular to a method and apparatus for fault diagnosis of satellite attitude control system based on time-frequency domain feature fusion. Background Technology

[0002] Fault diagnosis of satellites can ensure the success of critical space missions, improve system safety and reliability, and protect the near-Earth orbit environment. It is a key link in the aerospace field to ensure the long-term stable operation of satellites.

[0003] Currently, fault diagnosis methods for satellites typically include analytical model-based methods, expert knowledge-based methods, and data-driven methods. Due to the uncertainties of the on-orbit environment, analytical model-based methods struggle to provide a simple and highly accurate mathematical model. Expert knowledge-based methods require a large number of knowledge rules, making it difficult to form a complete rule base and lacking portability, thus resulting in low fault diagnosis efficiency. Data-driven methods suffer from the loss of detailed features in the original fault signal during feature extraction, affecting the diagnostic information and consequently leading to low accuracy. Summary of the Invention

[0004] In view of this, this disclosure aims to provide a method and apparatus for fault diagnosis of satellite attitude control system based on time-frequency domain feature fusion, which can improve the efficiency and accuracy of satellite fault diagnosis.

[0005] The technical solution disclosed herein is implemented as follows:

[0006] In a first aspect, this disclosure provides a fault diagnosis method for satellite attitude control systems based on time-frequency domain feature fusion, including:

[0007] Acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field;

[0008] Feature extraction is performed on the satellite image data to obtain a satellite feature image;

[0009] Extract global temporal features from satellite feature images, wherein the global temporal features refer to image sequence features extracted from the entire time series of the satellite that can represent the characteristics of the entire sequence;

[0010] The global temporal features are input into a dense connection model to obtain a fused temporal feature map corresponding to the satellite feature image;

[0011] The target fault type corresponding to the target satellite data is obtained based on the fused time series feature map.

[0012] Secondly, this disclosure provides a fault diagnosis device for a satellite attitude control system based on time-frequency domain feature fusion, the device comprising:

[0013] The data conversion module is used to acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field.

[0014] The feature extraction module extracts features from the satellite image data to obtain a satellite feature image;

[0015] The time series extraction module extracts global time series features from satellite feature images. The global time series features refer to image sequence features extracted from the entire time series of the satellite that can represent the characteristics of the entire sequence.

[0016] The feature fusion module inputs the global temporal features into a dense connection model to obtain a fused temporal feature map corresponding to the satellite feature image;

[0017] The fault classification module obtains the target fault type corresponding to the target satellite data based on the fused time series feature map.

[0018] Thirdly, this disclosure provides a computing device, the computing device comprising: a memory and a processor; wherein,

[0019] The memory is used to store computer programs that can run on the processor;

[0020] The processor is configured to execute the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion as described in the first aspect when running the computer program.

[0021] Fourthly, this disclosure provides a computer-readable storage medium storing at least one instruction, which is executed by a processor to implement the fault diagnosis method for satellite attitude control systems based on time-frequency domain feature fusion as described in the first aspect.

[0022] This disclosure provides a method and apparatus for fault diagnosis of satellite attitude control systems based on time-frequency domain feature fusion. By converting target satellite data into satellite image data and extracting features, the method automatically extracts global temporal features from the target satellite data to obtain global feature information characterizing the overall trend and pattern of the target satellite's operational status. Using image sequence features to represent satellite image data improves the expressive performance of features on satellite data, thereby enhancing the accuracy of fault diagnosis, reducing manual intervention, and improving the efficiency of fault diagnosis. The dense connection model design reduces the number of parameters, thus reducing computational resource consumption and improving the efficiency of model training and inference. Utilizing the dense connection structure of the dense connection model, deeper features are further extracted from the global temporal features, and features at different levels are fused, providing a more comprehensive description of the satellite's state. Furthermore, feature fusion reduces redundant computation, improves data processing efficiency, makes the fault diagnosis process faster, and enables the identification of more complex fault modes, adapting to different fault conditions and satellite environments. Based on the target fault type, the health and fault status of the satellite can be accurately diagnosed, improving the accuracy of fault diagnosis. Attached Figure Description

[0023] Figure 1 This disclosure provides a flowchart of a fault diagnosis method for satellite attitude control systems based on time-frequency domain feature fusion.

[0024] Figure 2 This is a schematic diagram of a network architecture based on feature fusion provided in this disclosure.

[0025] Figure 3 This is a schematic diagram of the composition of a GRU network model provided in this disclosure.

[0026] Figure 4 This is a flowchart of the training process for a GRU network model provided in this disclosure.

[0027] Figure 5 This is a schematic diagram of the composition of a dense connection model provided in this disclosure.

[0028] Figure 6 This is a schematic diagram of a confusion matrix for a fault detection result provided in this disclosure.

[0029] Figure 7 This is a schematic diagram illustrating the change curve of a loss function and training theory provided in this disclosure.

[0030] Figure 8 This is a schematic diagram illustrating the change curve of accuracy versus training discourse provided in this disclosure.

[0031] Figure 9This is a schematic diagram of a fault diagnosis device for a satellite attitude control system based on time-frequency domain feature fusion, as provided in this disclosure.

[0032] Figure 10 This is a schematic diagram of the hardware structure of a computing device provided in this disclosure. Detailed Implementation

[0033] The technical solutions in this disclosure will now be clearly and completely described with reference to the accompanying drawings.

[0034] Based on the above description of satellite fault diagnosis methods, it is evident that analytical model-based fault diagnosis methods struggle to provide a simple and highly accurate mathematical model. Expert knowledge-based fault diagnosis methods require a large number of knowledge rules, making it impossible to form a complete rule base and lacking portability, thus resulting in low fault diagnosis efficiency. Data-driven fault diagnosis methods suffer from the loss of detailed features in the original fault signal during feature extraction, affecting the fault diagnosis information and consequently leading to low accuracy. Therefore, this disclosure aims to provide a satellite attitude control system fault diagnosis technology solution based on time-frequency domain feature fusion, see [link to relevant documentation]. Figure 1 This disclosure illustrates a fault diagnosis method for a satellite attitude control system based on time-frequency domain feature fusion, which may include:

[0035] S101: Acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field.

[0036] Target satellite data typically originates from various sensors on the satellite, such as momentum wheel speed sensors, torque sensors, motor current and voltage sensors, and temperature sensors. These sensors collect various telemetry data during the satellite's operation at a certain sampling frequency, such as the momentum wheel's rotational speed, control torque, output torque, and engineering parameters like the satellite's attitude quaternions and angular velocity.

[0037] The target satellite data is time-series data, which has a time dimension. Different sensors may have different sampling frequencies, resulting in a large amount of data. It contains rich feature information that characterizes various working modes, states, and health conditions of components such as the satellite's momentum wheel. However, it also has problems such as difficulty in sampling alignment and a large number of irrelevant features.

[0038] The target satellite data is obtained by preprocessing the initial satellite data using a Kalman filter. The preprocessing includes at least normalization processing to normalize the initial satellite data to a uniform range, denoising processing to reduce noise in the initial satellite data, and segmentation processing to divide the initial satellite data into segments of fixed length according to a set time step to obtain the target satellite data, making the target satellite data suitable for input into the corresponding fault identification model.

[0039] In some examples, firstly, a certain type of data (such as momentum wheel rotation speed data) from the acquired satellite time series is treated as a time series of length N. Then, the cosine of the angle between each pair of data points in the time series is calculated, and an N×N Gram matrix is ​​constructed, where each element of the matrix represents the cosine of the angle between the corresponding two data points.

[0040] Specifically, suppose we have a set of vectors The Gram matrix is ​​derived from... The inner product matrix of each pair of vectors. As shown in the following equation, each element in the matrix... It is a vector and The vector product between them.

[0041]

[0042] Where G represents the Gram matrix, and the vectors in the Gram matrix are... and This represents the normalized scalar value at different time points in the time series. Specifically:

[0043] Given a time series of length N {x1, x2, ..., x...} N First, normalization is performed (scaled to [0,1] or [-1,1]), for each time point x. i Angles converted to polar coordinates =arccos(x i ). Gram matrix element G i , j =cos( + This reflects the temporal relationship between time points i and j. That is, a vector. and It is a normalized scalar time point value, rather than a traditional vector inner product.

[0044] The values ​​in the constructed Gram matrix are standardized, typically using a min-max standardization method to map them to the range [0,1]. This ensures data scaling, facilitating subsequent image representation and processing. The standardized Gram matrix is ​​then converted into a grayscale image. Specifically, each value in the matrix corresponds to the grayscale value of a pixel in the image, with the grayscale values ​​reflecting the relationships between time-series data points. Through this conversion, the original one-dimensional time-series satellite data is mapped to a two-dimensional image space, forming satellite image data.

[0045] S102: Perform feature extraction on satellite image data to obtain satellite feature images.

[0046] Satellite imagery may contain redundant information, and directly using these images for subsequent classification and recognition tasks can lead to noise interference and high computational costs. Feature extraction removes irrelevant and redundant information, highlighting features relevant to critical information such as satellite malfunctions, thereby improving the performance and efficiency of subsequent tasks. Feature extraction can be performed in various ways, such as using convolutional neural networks to extract features and obtain satellite feature images. Specific feature extraction methods will not be elaborated in this example implementation.

[0047] S103: Extract global temporal features from satellite feature images.

[0048] Global temporal features refer to image sequence features extracted from the entire time series of satellite data that represent the characteristics of the entire sequence. Typically, this involves collecting long-term satellite data—that is, identifying and extracting global sequence image sequence features from satellite feature images—that represent the trends and patterns of the entire dataset. These image sequence features usually include statistical parameters, frequency domain components, and temporal characteristics, capturing the overall changes in satellite operational status and providing crucial information for subsequent fault detection, health monitoring, and behavior prediction.

[0049] For example, the process of extracting global sequence features can be based on the satellite's momentum wheel, star sensor, and gyroscope components. The Gated Recurrent Unit (GRU) algorithm, Fourier transform, and autocorrelation analysis are used to extract global temporal features from the satellite feature images. The global temporal features are then filtered through principal component analysis and recursive feature elimination. The most representative features, namely star sensor quaternions, gyroscope rotation speed, and local rotation speed, are selected from all global temporal features to capture the overall pattern and trend information of satellite operation that is useful for fault detection and diagnosis.

[0050] S104: Input global temporal features into the dense connection model to obtain the fused temporal feature map corresponding to the satellite feature image.

[0051] Global temporal features can be input into the DenseNet model 250. In DenseNet, each layer is connected to all preceding layers, and the input of each layer is an ensemble of the outputs of all previous layers, allowing each layer to utilize the global feature information from all previous layers. Through dense connections between layers and the concatenation of global feature information along the channel dimension, feature fusion is achieved, resulting in a fused temporal feature map. This fused temporal feature map includes not only statistical information about the time series but also frequency domain features, temporal domain features, and nonlinear relationships; these features together constitute a multi-dimensional data representation.

[0052] S105: Obtain the target fault type corresponding to the target satellite data based on the fused time series feature map.

[0053] In some example embodiments of this disclosure, the fused timing feature map can be input into a Kalman filter for smoothing and correction to reduce the impact of noise.

[0054] For example, key features are extracted from the output of the Kalman filter and mapped onto a pre-trained fault identification model to obtain the target fault type and its location corresponding to the probability distribution of target satellite data under different fault types, and a corresponding fault report is generated.

[0055] According to the above scheme, by converting target satellite data into satellite image data and extracting features, and then automatically extracting global temporal features from the target satellite data, global feature information representing the overall trend and pattern of the target satellite's operational status is obtained. Using image sequence features to represent satellite image data improves the expressive performance of features on satellite data, thereby improving the accuracy of fault diagnosis, reducing manual intervention, and thus improving the efficiency of fault diagnosis. The design of the densely connected model 250 reduces the number of parameters, thereby reducing the consumption of computational resources and improving the training and inference efficiency of the model. Utilizing the densely connected structure of the densely connected model 250, deeper features are further extracted from the global temporal features, and features at different levels are fused, providing a more comprehensive description of the satellite's status. Furthermore, feature fusion reduces redundant computation, improves data processing efficiency, makes the fault diagnosis process faster, and enables the identification of more complex fault modes, adapting to different fault conditions and satellite environments. Based on the target fault type, the health and fault status of the satellite can be accurately diagnosed, improving the accuracy of fault diagnosis.

[0056] against Figure 1 The technical solution shown is described in the following reference. Figure 2 It illustrates a schematic diagram of a network architecture based on feature fusion provided in this disclosure, such as... Figure 2 As shown, the feature fusion-based network architecture may include a data preprocessing module 210, a signal conversion module 220, a feature extraction module 230, a GRU network model 240, and a dense connection model 250.

[0057] In this disclosure, the target satellite data is obtained by preprocessing the initial satellite data. Specifically, the data preprocessing module 210 described above can be used to preprocess the initial satellite data to obtain the target satellite data.

[0058] For example, initial satellite data can be acquired first, and then preprocessed using a Kalman filter to obtain target satellite data; the preprocessing includes at least normalization, denoising, and segmentation of the initial satellite data.

[0059] In the example above, the initial satellite data is preprocessed using a Kalman filter to obtain the target satellite data. Specifically, the preprocessing includes at least normalization, denoising, and segmentation of the initial satellite data. For example, data cleaning is performed to remove missing and outlier values ​​to ensure data quality; feature selection is performed to choose the features most relevant to the fault diagnosis task; normalization is implemented to ensure the data is within a uniform range, thereby improving the convergence speed and stability of model training; denoising is performed to reduce noise in the data, improving model accuracy; and time windowing is performed to divide the long-term time-series satellite data into multiple time windows, etc.

[0060] In some examples, a fixed-length time step, such as 24000, can be used as input to segment long-term satellite data into fixed-length sequences.

[0061] Based on the above example, after obtaining the target satellite data, the signal conversion module 220 can be used to convert the target satellite data to obtain satellite image data. Specifically, the Gram matrix can be used to obtain the satellite image data of the target satellite data. The acquisition process has been described in detail above, so it will not be repeated here.

[0062] After acquiring satellite image data, feature extraction module 230 can be used to extract features from the satellite image data to obtain satellite feature images. The specific feature extraction method has been explained above and will not be repeated here.

[0063] After obtaining satellite feature images, the GRU network model 240 can be used to extract features from the satellite feature images to obtain global temporal features.

[0064] Specifically, for the above example, various signal processing techniques can be used, such as GRU network model 240, Fourier transform, wavelet transform, and autocorrelation analysis, to extract global time series features from the target satellite data. The global time series features are then filtered through principal component analysis and recursive feature elimination, and the most representative features are selected from all extracted global time series features as the global time series features corresponding to the target satellite data.

[0065] See in some examples Figure 3Global temporal features of satellite feature images can be extracted using a GRU network model 240. The GRU network model 240 includes at least one feature extraction layer, at least one data processing layer, and a fully connected layer connected to the data processing layer for output. The process of extracting global temporal features from satellite feature images involves first inputting the preprocessed satellite feature image into the input layer of the GRU network model 240. The input and output of the feature extraction layer are then superimposed and input into the data processing layer. Finally, the input and output data of the data processing layer are superimposed and input into the next feature extraction layer or fully connected layer to obtain the global temporal features corresponding to the satellite feature image.

[0066] In the above example, specifically, the input layer of the GRU network model 240 receives preprocessed multi-dimensional satellite feature images and inputs the set batch size, time step (length of the sliding window), and feature number (input_feature_number) into the input layer. At least one feature extraction layer from a set of cascaded feature extraction layers and data processing layers is responsible for extracting key global temporal features from the satellite feature images input from the input layer. The feature extraction layer can be a convolutional layer or other types of layers used to capture patterns and trends in the satellite feature images.

[0067] In this disclosure, the feature extraction layer is a feature extraction layer in a recurrent neural network based on the GRU algorithm. The input and output of the feature extraction layer within this group are superimposed to form an enhanced feature representation, which is then input to the data processing layer. At least one data processing layer further processes the output of the feature extraction layer to enhance the nonlinear modeling capability of the features. The input and output data of the data processing layer are superimposed and input to the next feature extraction layer or directly to the fully connected layer. In this disclosure, when the input and output data of the data processing layer are superimposed and input to the next feature extraction layer, it is also necessary to reduce the feature dimensionality, extract key information, and reduce computational complexity. Then, the output of the data processing layer is input to the fully connected layer to obtain global temporal features.

[0068] Based on the aforementioned technical solution, in the specific implementation process, before extracting the global temporal features of satellite feature images, it is also necessary to train the initial network model based on the GRU algorithm using simulated sample feature images and real sample feature images to obtain the GRU network model 240. See also... Figure 4 This illustrates a training flowchart for a GRU network model 240 provided in this disclosure. In some examples, the training process may include:

[0069] S401: Obtain the sample feature image and the corresponding global temporal features;

[0070] S402: Use the feature extraction layer to extract features from the sample feature image to obtain the sample temporal features;

[0071] S403: Input the sample temporal features into the initial network model based on GRU to obtain reference temporal features;

[0072] S404: Update the initial network model and feature extraction layer based on the real global temporal features and reference temporal features to obtain the GRU network model.

[0073] For the example above, satellite-based sample feature images are labeled with corresponding global temporal features, and each temporal feature is assigned a unique label or code for subsequent model training. In one exemplary embodiment of this disclosure, the GRU network model 240 is primarily based on a deep learning neural network model. For example, the GRU network model 240 can be based on a feedforward neural network. The feedforward network can be implemented as an acyclic graph, where nodes are arranged in layers. Typically, the feedforward network topology includes an input layer and an output layer, which are separated by at least one hidden layer. The hidden layer transforms the input received by the input layer into a representation useful for generating the output in the output layer. Network nodes are fully connected to nodes in adjacent layers via edges, but there are no edges between nodes within each layer. Data received at the nodes of the input layer of the feedforward network is propagated (i.e., “feedforward”) to the nodes of the output layer via an activation function, which calculates the state of the nodes in each consecutive layer of the network based on coefficients (“weights”), which are associated with each of the edges connecting these layers. The output of the GRU network model 240 can take various forms, and this disclosure does not limit this. The GRU network model 240 can also be other neural network models, such as convolutional neural network (CNN) models, recurrent neural network (RNN) models, generative adversarial network (GAN) models, but is not limited thereto, and can also employ other neural network models known to those skilled in the art.

[0074] Specifically, the training process described above involves first acquiring corresponding sample feature images, and then training the initial network model using these sample feature images. This can include the following steps: selecting a network topology; using a set of training data representing the problem being modeled by the network; and adjusting the weights until the network model exhibits minimum error for all instances in the training dataset. For example, during supervised learning training for a neural network, the output generated by the network in response to inputs representing instances in the training dataset is compared to the "correct" labeled output of that instance; an error signal representing the difference between the output and the labeled output is calculated; and the weights associated with the connections are adjusted to minimize this error as the error signal is backpropagated through the network layers. The model where the error of each output generated from the instances in the training dataset is minimized is defined as the GRU network model 240.

[0075] Based on the foregoing explanation, after obtaining the global temporal features, feature fusion is still required using a densely connected model 250. For example, depending on the importance of the features or a specific fusion strategy, in some possible implementations, at least two intermediate features can be selected for feature fusion. Alternatively, the global temporal features can be fused with at least one of multiple intermediate features, which can come from different layers to capture multi-scale information. Specific feature fusion methods can be selected based on prior knowledge, feature importance assessment, or experimental results. For example, feature concatenation, weighted summation, or more complex fusion strategies can be used to integrate the information of the selected features. Feature concatenation can involve concatenating selected features along the channel dimension to achieve feature fusion, increasing the feature dimensionality and information content.

[0076] In some examples, feature fusion via densely connected model 250 first requires constructing the network structure of densely connected model 250, such as... Figure 5 As shown, the dense connection model 250 consists of multiple sequentially connected feature fusion layers, with each layer connected to the previous one, ensuring the flow of information between layers and the accumulation of features.

[0077] In some examples, such as Figure 5 As shown, the feature fusion layer includes multiple convolutional modules; inputting global temporal features into the first feature fusion layer to obtain the first fused temporal feature map may include: using multiple convolutional modules to perform convolution processing on the global temporal features to obtain multiple intermediate features; wherein, the input of the P-th convolutional module is the output of the (P-1)-th convolutional module, and P is greater than or equal to 2; fusing at least two of the multiple intermediate features to obtain the first fused temporal feature map.

[0078] In the example above, specifically, the feature fusion layer consists of multiple convolutional modules, each responsible for extracting features at different levels. Each convolutional module can include settings for parameters such as kernel size, stride, and padding to adapt to the extraction needs of different features. Global temporal features are transmitted as input data to the first-layer convolutional module. Optionally, the first-layer convolutional module performs convolution operations on the input global temporal features to generate the first-layer convolutional temporal features. The convolution operation is performed in each feature fusion layer, using convolutional kernels to extract local features to obtain the intermediate features of the corresponding layer. Optionally, the first-layer convolutional module performs batch normalization and non-linear activation functions on the input global temporal features. Batch normalization is applied to reduce internal covariate shifts; non-linear activation introduces non-linearity through non-linear activation functions such as ReLU to enhance the model's expressive power. The output of each convolutional module serves as the input to the next convolutional module, forming multiple layers of intermediate features. For the P-th convolutional module (where P ≥ 2), its input is the output of the (P-1)-th convolutional module, ensuring information flow between layers. In multi-layered intermediate features, before fusion, feature selection techniques, such as those based on the coefficient of variation or information gain, can be applied to select at least two intermediate features with the most information for fusion.

[0079] After the fusion operation, the first fused temporal feature map is obtained, namely the fused temporal feature map of feature fusion layer 1. This fused temporal feature map integrates the information extracted by multiple convolutional modules. The first fused temporal feature map will be used as the input of the second fused feature layer. Correspondingly, the output of the second fused feature layer, namely feature fusion layer 2, will be used as the input of subsequent layers or for further processing of the fault diagnosis model.

[0080] Through the aforementioned multi-layer convolutional processing and feature fusion, the first feature fusion layer can effectively extract and integrate key information from global temporal features, providing strong data support for fault diagnosis. Fusing intermediate features from different levels can capture multi-scale information, improving the accuracy of fault diagnosis.

[0081] For example, see Figure 5 Assuming the densely connected model 250 includes M sequentially connected feature fusion layers, inputting global temporal features into the densely connected model 250 to obtain the fused temporal feature map corresponding to the satellite feature image can include: inputting global temporal features into the first feature fusion layer to obtain the first fused temporal feature map; using the output of the Nth feature fusion layer as the input of the N+1th feature fusion layer to obtain the N+1th fused temporal feature map; and using the fused temporal feature map of the Mth layer as the target fused temporal feature map.

[0082] For the above example, specifically, the extracted global temporal features are transmitted as input data to the first feature fusion layer of the densely connected model 250. In the first feature fusion layer, the input global temporal features are preliminarily processed, such as the convolution operation shown in the foregoing example content. The first fused temporal feature map is used as the input for the next layer for feature transmission. Each subsequent feature fusion layer receives the outputs from all previous layers as input to achieve feature accumulation and feature fusion. For the feature fusion layer of the Nth layer (1 < N < M), the operations of convolution, normalization, and activation function are repeatedly executed, and the output map. The output is passed to the (N + 1)th layer, which will not be elaborated here in detail. After the Mth layer, that is, the last feature fusion layer, finishes processing, the target fused temporal feature map is output. The target fused temporal feature map contains all the information accumulated from the input to the last layer. Through the method of layer-by-layer feature fusion, the densely connected model 250 can effectively extract and integrate key information from the global temporal features, providing strong data support for fault diagnosis.

[0083] In some examples, when obtaining the target fused temporal feature map, at least one of the global temporal features and multiple intermediate features can be fused to obtain the fused temporal feature map corresponding to the satellite feature image.

[0084] Exemplarily, through the above satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion, the global temporal features can also be fused with at least one intermediate feature. First, they are concatenated in the channel dimension to increase the feature dimension. Secondly, weighted summation is performed, and the weights can be dynamically adjusted according to the importance of the features. Then, element-wise multiplication or division operations are executed to combine the global temporal features and the intermediate features to achieve feature fusion. The features after the fusion operation and processing are used as the fused temporal feature map corresponding to the satellite feature image, and this fused temporal feature map synthesizes the information of the global temporal features and the intermediate features.

[0085] It should be noted that based on the foregoing training process of the GRU network model 240, correspondingly, the densely connected model 250 also needs to be trained. Comparing with the training process of the GRU network model 240, the input of the GRU network model 240 is the target satellite data, and the output is the global temporal features. The input of the densely connected model 250 is the global temporal features obtained through the GRU network model 240, and the output is the fused temporal feature map. For the specific training process, please refer to the elaboration of the foregoing training process content, which will not be elaborated here in detail.

[0086] Based on the foregoing explanation, after obtaining the fused temporal feature map, a Kalman filter is constructed, and its parameters are defined, such as the state transition matrix, observation matrix, process noise covariance matrix, and observation noise covariance matrix. The outputs of the GRU network model 240 and the densely connected model 250, namely the global temporal features and the fused temporal feature map, are input into the Kalman filter. The Kalman filter smooths and corrects the output results, reducing the influence of random noise and correcting the estimation of the system state, thereby providing more accurate and stable fault diagnosis results.

[0087] In some examples, obtaining the target fault type corresponding to the satellite feature image based on the fused temporal feature map may include: obtaining the fault representation vector corresponding to the satellite feature image based on the fused temporal feature map; and determining the target fault type corresponding to the satellite feature image based on the fault representation vector.

[0088] Specifically, in the example above, taking the fused temporal feature map of target satellite data as input into a pre-trained fault identification model, the pre-trained fault identification model can include a global average pooling layer, a fully connected layer, and an output layer. The hierarchical processing in the pre-trained fault identification model captures complex patterns and trends in the target satellite data. The global average pooling layer reduces feature dimensionality and extracts key information, ultimately outputting a fault representation vector. This fault representation vector contains a high-level feature representation of the input data. Based on the fault representation vector, the pre-trained fault identification model determines the fault type corresponding to the target satellite data through the classification decision mechanism of the fully connected layer. This process involves mapping the fault representation vector to a probability distribution, comparing the probability distribution values ​​with set thresholds to obtain the output values ​​corresponding to the fault types, and combining the output values ​​of the fault types to obtain the fault representation vector.

[0089] In some examples, the fault type with the highest probability distribution value can also be identified as the target fault type corresponding to the satellite feature image. This approach not only improves the accuracy and efficiency of fault diagnosis, but also reduces the reliance on large amounts of labeled data by applying pre-trained fault identification models, making the fault identification process more efficient and practical.

[0090] Based on the above example, obtaining the fault representation vector corresponding to the satellite feature image from the fused temporal feature map may include: determining at least one candidate fault type; determining the probability distribution value of the target satellite data for each candidate fault type based on the fused temporal feature map; and obtaining the fault representation vector based on the probability distribution value.

[0091] For the example above, one or more candidate fault types are first identified. Then, based on one or more candidate fault types, a corresponding probability distribution value is calculated for each candidate fault type. Finally, the probability distribution values ​​are combined to form a fault characterization vector, thereby providing a comprehensive fault type probability description for satellite data. The fault characterization vector summarizes the probability that satellite data belongs to each fault type.

[0092] For example, taking a momentum wheel as an example, the above method can be used to identify six types of momentum wheel malfunctions (idling malfunction, jamming malfunction, excessive friction malfunction, abnormal noise malfunction, and periodic oscillation malfunction) and normal conditions, resulting in the following: Figure 6 The confusion matrix shown indicates that the larger the value, the greater the predicted probability. The maximum value in each row represents the maximum probability. The conclusion obtained is consistent with the true value, that is, the method disclosed in this paper can accurately determine the fault type.

[0093] Reference Figure 7 and Figure 8 During the training process of the fault identification model, as the number of training rounds increases, the loss function gradually converges and the accuracy gradually improves.

[0094] In some examples, the fault identification model can be a GASF-DenseNet-GRU model, a Long Short-Term Memory network model, a convolutional neural network model, or a GRU network model, etc. See Table 1 for details:

[0095] Table 1 Comparison of Network Results

[0096]

[0097] As shown in Table 1, the GASF-DenseNet-GRU fault identification model achieved a 92% accuracy rate in classifying five types of faults in the satellite attitude control subsystem.

[0098] This fault signal classifier has the best fault identification effect for momentum wheel jamming and excessive friction faults.

[0099] This disclosure combines the GRU network model with a densely connected model to provide more comprehensive data analysis. The GRU network model processes the entire long-sequence data of the satellite to extract global temporal features from the target satellite data, namely global statistical features, frequency domain features, and temporal domain features. Then, the densely connected model fuses features of different scales and types, giving the model better generalization ability and adaptability to different failure modes and environmental changes. Furthermore, the densely connected model reduces the number of parameters through feature reuse, improving data processing efficiency.

[0100] Based on the same inventive concept as the aforementioned technical solution, see [link to inventive concept].Figure 9 This disclosure illustrates a fault diagnosis device 900 for a satellite attitude control system based on time-frequency domain feature fusion. The device 900 may include: a data conversion module 901, a feature extraction module 902, a time-series extraction module 903, a feature fusion module 904, and a fault classification module 905.

[0101] The data conversion module 901 can be used to acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field;

[0102] The feature extraction module 902 can be used to extract features from the satellite image data to obtain satellite feature images;

[0103] The time series extraction module 903 can be used to extract global time series features of satellite feature images, wherein the global time series features refer to image sequence features extracted from the entire time series of the satellite that can represent the characteristics of the entire sequence.

[0104] The feature fusion module 904 can be used to input the global temporal features into a dense connection model to obtain the fused temporal feature map corresponding to the satellite feature image;

[0105] The fault classification module 905 can be used to obtain the target fault type corresponding to the target satellite data based on the fused time series feature map.

[0106] In some examples, feature fusion module 904 is configured as follows:

[0107] The global temporal features are input into the first feature fusion layer to obtain the first fused temporal feature map;

[0108] The output of the Nth feature fusion layer is used as the input of the N+1th feature fusion layer to obtain the fused temporal feature map of the N+1th layer, where N is greater than 1 and less than M-1;

[0109] The fusion temporal feature map of the Mth layer is used as the target fusion temporal feature map.

[0110] In some examples, feature fusion module 904 is configured as follows:

[0111] Multi-layer convolutional modules are used to perform convolutional processing on global temporal features to obtain multi-layer intermediate features; wherein the input of the P-th convolutional module is the output of the (P-1)-th convolutional module, and P is greater than or equal to 2;

[0112] At least two of the intermediate features from the multi-layered system are fused to obtain the first-layer fused temporal feature map.

[0113] In some examples, feature fusion module 904 is configured as follows:

[0114] The fused temporal feature map corresponding to the satellite feature image is obtained by fusing global temporal features and at least one of multiple intermediate features.

[0115] In some examples, the timing extraction module 903 is configured as follows:

[0116] Acquire initial satellite data;

[0117] The initial satellite data is preprocessed using a Kalman filter to obtain the target satellite data. The preprocessing includes at least normalization, denoising, and segmentation of the initial satellite data to obtain the target satellite data.

[0118] In some examples, the timing extraction module 903 is configured as follows:

[0119] A GRU network model is used to extract features from target satellite data to obtain global temporal features; among which,

[0120] The GRU network model includes at least one feature extraction layer, at least one data processing layer, and a fully connected layer for output connected to the data processing layer, all of which are individually connected to each other.

[0121] The input and output of the feature extraction layer are superimposed and fed into the data processing layer;

[0122] The input and output data of the data processing layer are superimposed and fed into the next feature extraction layer or fully connected layer to obtain the global temporal features corresponding to the target satellite data.

[0123] In some examples, the fault classification module 905 is configured as follows:

[0124] Obtain the fault characterization vector corresponding to the target satellite data based on the fused temporal feature map;

[0125] The target fault type corresponding to the target satellite data is determined based on the fault characterization vector.

[0126] It is understood that the exemplary technical solution of the satellite attitude control system fault diagnosis device 900 based on time-frequency domain feature fusion described above belongs to the same concept as the technical solution of the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion described above. Therefore, all details not described in detail in the technical solution of the satellite attitude control system fault diagnosis device 900 based on time-frequency domain feature fusion described above can be found in the description of the technical solution of the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion described above. This disclosure will not elaborate further on these details.

[0127] Please refer to Figure 10This illustration shows a schematic diagram of the hardware structure of a computing device provided in an exemplary embodiment of this disclosure. In some examples, the computing device can be at least one of devices such as a smartphone, smartwatch, desktop computer, laptop, virtual reality terminal, augmented reality terminal, wireless terminal, and laptop computer. The computing device has communication functions and can access wired or wireless networks. The computing device can refer to one of multiple terminals; those skilled in the art will understand that the number of such terminals can be more or less. In some examples, the computing device can receive data from a satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion, based on the accessed wired or wireless network. It is understood that the computing device undertakes the computation and processing work of the technical solution of this disclosure, and this disclosure does not limit it in this regard.

[0128] like Figure 10 As shown, the computing device in this disclosure may include one or more of the following components: processor 1010 and memory 1020.

[0129] Optionally, the processor 1010 connects various parts within the computing device using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1020, and by calling data stored in the memory 1020. Optionally, the processor 1010 can be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 1010 can integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), Neural-network Processing Unit (NPU), and baseband chip. Specifically, the CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content displayed on the touchscreen; the NPU implements Artificial Intelligence (AI) functions; and the baseband chip handles wireless communication. It is understandable that the aforementioned baseband chip may not be integrated into the processor 1010, but may be implemented using a separate chip.

[0130] The memory 1020 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 1020 may include a non-transitory computer-readable storage medium. The memory 1020 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 1020 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data created according to the use of the computing device, etc.

[0131] In addition, those skilled in the art will understand that the structure of the computing device shown in the above figures does not constitute a limitation on the computing device. The computing device may include more or fewer components than shown, or combine certain components, or have different component arrangements. For example, the computing device may also include a display screen, camera assembly, microphone, speaker, radio frequency circuit, input unit, sensors (such as accelerometer, angular velocity sensor, fiber optic sensor, etc.), audio circuit, WiFi module, power supply, Bluetooth module, etc., which will not be described in detail here.

[0132] This disclosure also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor to implement the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion as described in the various embodiments above.

[0133] This disclosure also provides a computer program product including computer instructions stored in a computer-readable storage medium; a processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to perform the fault diagnosis method for a satellite attitude control system based on time-frequency domain feature fusion as described in the various embodiments above.

[0134] Those skilled in the art will recognize that the functions described in this disclosure in one or more of the foregoing examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer-readable storage media and communication media, wherein communication media include any medium that facilitates the transfer of a computer program from one place to another. A readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0135] The above are merely specific embodiments of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A fault diagnosis method for satellite attitude control systems based on time-frequency domain feature fusion, characterized in that, The method includes: Acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field, wherein the target satellite data is telemetry data collected by sensors during satellite operation and the target satellite data is time series data; Feature extraction is performed on the satellite image data to obtain a satellite feature image; The GRU network model is used to extract features from the satellite feature images to obtain global temporal features, wherein the global temporal features refer to image sequence features extracted from the entire time series of the satellite that can represent the characteristics of the entire sequence. The global temporal features are input into a dense connection model to obtain a fused temporal feature map corresponding to the satellite feature image, wherein the dense connection model includes M sequentially connected feature fusion layers; The fault characterization vector corresponding to the target satellite data is obtained based on the fused temporal feature map, and the target fault type corresponding to the target satellite data is determined based on the fault characterization vector. The step of inputting the global temporal features into a dense connection model to obtain the fused temporal feature map corresponding to the satellite feature image includes: The global temporal features are input into the first feature fusion layer to obtain the first fused temporal feature map; The output of the Nth feature fusion layer is used as the input of the N+1th feature fusion layer to obtain the fused temporal feature map of the N+1th layer, where N is greater than 1 and less than M-1; The fused temporal feature map of the Mth layer is used as the target fused temporal feature map.

2. The method according to claim 1, characterized in that, The GRU network model includes at least one feature extraction layer, at least one data processing layer, and a fully connected layer for output connected to the data processing layer, all of which are individually connected to each other.

3. The method according to claim 2, characterized in that, The step of using a GRU network model to extract features from the satellite feature image to obtain global temporal features includes: The input and output of the feature extraction layer are superimposed and input into the data processing layer; The input and output data of the data processing layer are superimposed and input into the next feature extraction layer or fully connected layer to obtain the global temporal features corresponding to the target satellite data.

4. The method according to claim 1, characterized in that, The feature fusion layer includes multiple convolutional modules; the step of inputting the global temporal features into the first feature fusion layer to obtain the first fused temporal feature map includes: The global temporal features are processed by the multi-layer convolutional module to obtain multi-layer intermediate features; wherein the input of the P-th layer convolutional module is the output of the (P-1)-th layer convolutional module, and P is greater than or equal to 2; At least two of the intermediate features from the multi-layered system are fused to obtain the first-layer fused temporal feature map.

5. The method according to claim 4, characterized in that, The step of inputting the global temporal features into a dense connection model to obtain the fused temporal feature map corresponding to the satellite feature image includes: The global temporal features and at least one of multiple intermediate features are fused to obtain the fused temporal feature map corresponding to the satellite feature image.

6. The method according to claim 1, characterized in that, The method further includes: Acquire initial satellite data; The initial satellite data is preprocessed using a Kalman filter to obtain the target satellite data. The preprocessing includes at least normalizing, denoising, and segmenting the initial satellite data to obtain the target satellite data.

7. The method according to claim 1, characterized in that, The sensor includes at least one of the following: momentum wheel speed sensor, torque sensor, motor current and voltage sensor, and temperature sensor.

8. A fault diagnosis device for a satellite attitude control system based on time-frequency domain feature fusion, characterized in that, The device includes: The data conversion module is used to acquire target satellite data and convert the target satellite data into satellite image data using Gram angle field. The target satellite data is telemetry data collected by sensors during satellite operation and is time-series data. The feature extraction module is used to extract features from the satellite image data to obtain satellite feature images; The time series extraction module is used to extract features from the satellite feature images using a GRU network model to obtain global time series features, wherein the global time series features refer to image sequence features extracted from the entire time series of the satellite that can represent the characteristics of the entire sequence. The feature fusion module is used to input the global temporal features into a dense connection model to obtain a fused temporal feature map corresponding to the satellite feature image, wherein the dense connection model includes M sequentially connected feature fusion layers; Fault classification module, used for The fault characterization vector corresponding to the target satellite data is obtained based on the fused temporal feature map, and the target fault type corresponding to the target satellite data is determined based on the fault characterization vector. The step of inputting the global temporal features into a dense connection model to obtain the fused temporal feature map corresponding to the satellite feature image includes: The global temporal features are input into the first feature fusion layer to obtain the first fused temporal feature map; The output of the Nth feature fusion layer is used as the input of the N+1th feature fusion layer to obtain the fused temporal feature map of the N+1th layer, where N is greater than 1 and less than M-1; The fused temporal feature map of the Mth layer is used as the target fused temporal feature map.

9. A computing device, characterized in that, The computing device includes: a processor and a memory; wherein... The memory is used to store computer programs that can run on the processor; The processor is configured to execute, when running the computer program, the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores at least one instruction, which is executed by a processor to implement the satellite attitude control system fault diagnosis method based on time-frequency domain feature fusion as described in any one of claims 1 to 7.

Images