Satellite measurement and control data anomaly detection and intelligent diagnosis early warning method and system

Abstract

The invention provides a satellite measurement and control data anomaly detection and intelligent diagnosis early warning method and system, and relates to the technical field of satellite measurement and control, and the method comprises the steps: obtaining multi-source heterogeneous data in a satellite measurement and control system, and carrying out the preprocessing of the multi-source heterogeneous data; constructing a multi-level space-time fusion network model, inputting the preprocessed multi-source heterogeneous data into the multi-level space-time fusion network model, and extracting high-dimensional space-time features; based on the extracted high-dimensional spatial-temporal features, designing a dynamic adaptive anomaly detection algorithm, and performing anomaly detection on the satellite measurement and control data; according to an anomaly detection result, performing intelligent diagnosis by using a pre-constructed multi-level satellite system knowledge graph; based on the intelligent diagnosis result, multi-level early warning information is generated; a distributed edge computing architecture is designed, and partial data processing and anomaly detection tasks are deployed to a ground station and a satellite end; and presenting the generated multi-level early warning information through a visual interface, and automatically triggering a corresponding emergency plan according to the early warning level.

Description

Technical Field

[0001] The present invention relates to satellite measurement and control technology, and in particular to a satellite measurement and control data anomaly detection and intelligent diagnosis and early warning method and system. Background Art

[0002] Existing satellite measurement and control data processing methods mainly rely on expert experience and fixed rules, and use simple statistical analysis and threshold judgment for anomaly detection, which is difficult to adapt to the complex and changing space environment.

[0003] Traditional anomaly detection methods cannot effectively handle the correlation between multi-source heterogeneous data such as telemetry data, orbit data, and attitude data, resulting in insufficient reliability of detection results. At the same time, since satellites are affected by various factors such as space environment, sunlight conditions, and mission mode during their in-orbit operation, measurement and control data show obvious time-varying characteristics and non-stationarity, and fixed detection rules and thresholds are difficult to meet actual needs. In terms of fault diagnosis, most existing methods use a simple expert rule base, lacking in-depth analysis of the system hierarchy and the correlation between components, making it difficult to quickly locate faults and track propagation paths.

[0004] In addition, as satellite systems become increasingly complex and networked, the amount of measurement and control data generated is growing exponentially, and traditional centralized computing architectures face severe challenges in terms of real-time and scalability. The computing resources of ground stations and satellites are not fully utilized, and data transmission and processing delays significantly affect the efficiency of anomaly detection and fault diagnosis. The existing early warning mechanism is mainly based on simple threshold judgments, lacks the ability to fine-tune the risk level and dynamically adjust it, and is difficult to provide effective decision support for emergency response. Summary of the invention

[0005] The embodiments of the present invention provide a satellite measurement and control data anomaly detection and intelligent diagnosis and early warning method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provide satellite measurement and control data anomaly detection and intelligent diagnosis and early warning methods, including:

[0008] Acquire multi-source heterogeneous data in a satellite measurement and control system, wherein the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data, and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization, and time alignment; construct a multi-level spatiotemporal fusion network model, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data, and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; design a dynamic adaptive anomaly detection algorithm based on the extracted high-dimensional spatiotemporal features to perform anomaly detection on satellite measurement and control data;

[0009] Based on the anomaly detection results, intelligent diagnosis is performed using the pre-built multi-level satellite system knowledge graph. Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm to generate multi-level warning information. A distributed edge computing architecture is designed to deploy some data processing and anomaly detection tasks to ground stations and satellite terminals.

[0010] Build a human-machine collaborative decision-making system to present the generated multi-level warning information through a visual interface, and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on operator feedback information and historical data.

[0011] In an optional embodiment,

[0012] A multi-level spatiotemporal fusion network model is constructed, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data, and an attention mechanism layer for feature fusion; the pre-processed multi-source heterogeneous data is input into the multi-level spatiotemporal fusion network model, and the steps of extracting high-dimensional spatiotemporal features include:

[0013] The bidirectional gated recurrent unit layer processes past and future context information simultaneously through forward and backward gated recurrent units to capture the long-term dependencies of satellite states. The graph convolutional network layer represents the connection relationship between the components of the satellite system based on a predefined adjacency matrix and processes non-Euclidean spatial data through graph convolution operations. The attention mechanism layer uses a multi-head self-attention mechanism to learn the importance weights between different features.

[0014] The pre-processed multi-source heterogeneous satellite measurement and control data are input into the multi-level spatiotemporal fusion network model, where the time series data is input into the bidirectional gated recurrent unit layer, and the spatial data is input into the graph convolutional network layer;

[0015] Use the bidirectional gated recurrent unit layer to process the time series data to obtain the time series features; use the graph convolutional network layer to process the spatial data to obtain the spatial features;

[0016] The obtained temporal features and spatial features are concatenated, including: concatenating the temporal feature matrix and the spatial feature matrix on the feature dimension to form a comprehensive feature matrix, in which each row represents a time step or a spatial node, and each column represents a feature dimension; the concatenated temporal features and spatial features are input into the attention mechanism layer, and the fused features are obtained through multi-head self-attention calculation.

[0017] In an optional embodiment,

[0018] The method further comprises:

[0019] Constructing a multi-layer perceptron structure, wherein the multi-layer perceptron structure includes three fully connected layers, each of which is followed by a nonlinear activation function;

[0020] The fused features are input into the first fully connected layer of the multi-layer perceptron structure, the input fused features are linearly transformed, and nonlinearly mapped through a nonlinear activation function to obtain the first layer output features; the first layer output features are input into the second fully connected layer, and a second linear transformation and nonlinear mapping are performed to obtain the second layer output features; a Dropout layer is added between the second layer output features and the third fully connected layer; the second layer output features processed by Dropout are input into the third fully connected layer, and a final linear transformation is performed to obtain high-dimensional spatiotemporal features; wherein the nonlinear activation function adopts a rectified linear unit function.

[0021] In an optional embodiment,

[0022] Based on the extracted high-dimensional spatiotemporal features, a dynamic adaptive anomaly detection algorithm is designed. The steps of anomaly detection on satellite measurement and control data include:

[0023] A dynamic adaptive anomaly detection algorithm is designed by combining a mixed Gaussian model with a dynamic threshold; the mixed Gaussian model models the probability distribution of high-dimensional spatiotemporal features, and the parameters of the mixed Gaussian model are estimated by an expectation maximization algorithm. The mixed Gaussian model includes multiple Gaussian components, each of which has a corresponding mixing coefficient, a mean vector, and a covariance matrix;

[0024] Calculating the negative log-likelihood score of the high-dimensional spatiotemporal features using the mixed Gaussian model, setting a time window, calculating the mean and standard deviation of the negative log-likelihood score within the time window, and constructing a dynamic threshold based on the mean and standard deviation;

[0025] Perform an exponentially weighted moving average on the negative log-likelihood scores of multiple consecutive time steps to obtain a smoothed score sequence; calculate the Mahalanobis distance of the high-dimensional spatiotemporal features, where the Mahalanobis distance is calculated based on the mean vector and the covariance matrix; perform a weighted combination of the smoothed score sequence and the Mahalanobis distance to construct a composite anomaly metric; compare the composite anomaly metric with a dynamic threshold, and when the composite anomaly metric exceeds the dynamic threshold, mark the current time step as an anomaly point; analyze the contribution of the high-dimensional spatiotemporal features marked as anomalies in each feature dimension, and identify the feature dimensions that cause the anomaly.

[0026] In an optional embodiment,

[0027] Based on the anomaly detection results, the steps for intelligent diagnosis using the pre-built multi-level satellite system knowledge graph include:

[0028] Construct a multi-level satellite system knowledge graph containing physical device nodes, functional module nodes, state parameter nodes, fault mode nodes and symptom feature nodes, and establish physical inclusion relationships, functional dependency relationships, parameter association relationships, fault propagation relationships and symptom attribution relationships between nodes;

[0029] The multi-level satellite system knowledge graph is modeled by using a graph attention network, the attention weights between nodes are calculated through a multi-head attention mechanism, and the feature representation of the nodes is updated according to the attention weights; the anomaly detection results are input into the graph attention network, the corresponding symptom feature nodes are activated, and the activation strength of the symptom nodes is obtained; based on the activation strength of the symptom nodes, the fault propagation process is simulated in the multi-level satellite system knowledge graph, and the activation state of each fault node is calculated; according to the activation state of the fault node, the root cause of the fault is determined by reverse reasoning, and the association strength of the root cause node is obtained; according to the association strength of the root cause node, the path score connecting the symptom node and the fault node is calculated, and the path above the importance threshold is extracted to form a diagnostic subgraph;

[0030] Based on the diagnosis subgraph, evidence sources are extracted from three dimensions: physical equipment, functional modules and state parameters, and probability weights are assigned to each evidence source according to the importance of each dimension. The Dempster combination rule is used to fuse the evidence sources to obtain the final diagnosis result.

[0031] In an optional embodiment,

[0032] Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm. The steps to generate multi-level warning information include:

[0033] Acquire real-time diagnostic results and system operation data, convert the real-time diagnostic results into binary feature vectors through one-hot encoding, fuse the system operation data with the binary feature vectors after standardization to obtain fused features; establish an action space including risk level assessment and warning level selection;

[0034] Construct a dual deep Q network structure, including a main network and a target network. Both the main network and the target network are composed of a multi-channel input layer, a convolutional feature extraction layer, a fully connected fusion layer, and an action value output layer.

[0035] Calculate the loss function including the temporal difference error, use the stochastic gradient descent method to update the main network parameters, and periodically synchronize the main network parameters to the target network through the soft update strategy;

[0036] Based on the fusion features, the Q value of each action is calculated through the main network, and the evaluation action is selected to determine the risk level and warning level, generating multi-level warning information including risk determination basis, potential hazard analysis and disposal suggestions.

[0037] In an optional embodiment,

[0038] The steps to design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to ground stations and satellites include:

[0039] A ground station distributed computing framework including a data preprocessing layer, a model training layer, and an inference service layer is constructed. The data preprocessing layer distinguishes duplicate data through a Bloom filter and performs distributed processing of data cleaning and feature engineering with an incremental storage mechanism to generate a standardized training data set; the model training layer uses a parameter server architecture and a ring full reduction algorithm to perform model parallel training based on the training data set, and migrates the trained complex model knowledge to a lightweight model through a model distillation method, which is deployed to the inference service layer for online inference; the inference service layer monitors and counts the computing load indicator information of each computing node in real time, and the computing load indicator information includes CPU usage, memory occupancy, and task queue length;

[0040] Establish an edge node task scheduling mechanism, receive the computing load indicator information of the inference service layer, and allocate new tasks according to the task complexity, network quality and node computing power. When the task complexity is lower than the preset threshold, it is allocated to the satellite end; when the network quality is lower than the preset threshold, it is allocated to the nearest ground station.

[0041] Perform data compression and incremental update processing on the assigned tasks, compress the transmitted data according to the adaptive compression ratio, and obtain the status of each node through the lightweight heartbeat protocol; generate checkpoints and snapshot backups during task execution based on the node status, and read the task status from the most recent checkpoint data when a node failure is detected, and migrate the task to other idle nodes;

[0042] The task data is stored in a multi-copy mechanism, and the task execution results are transmitted to the data preprocessing layer in an incremental update manner for updating the training data set.

[0043] According to a second aspect of the embodiments of the present invention,

[0044] Provide satellite measurement and control data anomaly detection and intelligent diagnosis and early warning system, including:

[0045] The first unit is used to obtain multi-source heterogeneous data in the satellite measurement and control system, and the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization and time alignment; construct a multi-level spatiotemporal fusion network model, and the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; based on the extracted high-dimensional spatiotemporal features, design a dynamic adaptive anomaly detection algorithm to perform anomaly detection on satellite measurement and control data;

[0046] The second unit is used to perform intelligent diagnosis based on the anomaly detection results using the pre-built multi-level satellite system knowledge graph; based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm to generate multi-level warning information; a distributed edge computing architecture is designed to deploy some data processing and anomaly detection tasks to ground stations and satellite terminals;

[0047] The third unit is used to build a human-machine collaborative decision-making system, which will present the generated multi-level warning information through a visual interface and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on the operator's feedback information and historical data.

[0048] According to a third aspect of the embodiments of the present invention,

[0049] An electronic device is provided, comprising:

[0050] processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] According to a fourth aspect of the embodiments of the present invention,

[0054] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0055] The satellite measurement and control data anomaly detection and intelligent diagnosis and early warning system provided by the present invention has significant technical effects. By constructing a multi-level spatiotemporal fusion network model, the effective fusion of satellite telemetry data, orbit data, attitude data and ground station equipment status data is achieved, and the temporal correlation and spatial dependency between different types of data are fully explored, which significantly improves the accuracy and reliability of anomaly detection.

[0056] Based on the design of dynamic adaptive anomaly detection algorithm, the system can automatically adjust the detection strategy and parameter threshold according to the actual operation status of the satellite and the changes in the space environment, effectively reducing the false alarm rate and missed alarm rate, and enhancing the credibility of the detection results. At the same time, the use of multi-level satellite system knowledge graph for intelligent diagnosis can not only quickly locate the cause of the fault, but also analyze the propagation path and impact range of the fault, providing strong support for fault handling.

[0057] The present invention uses a deep reinforcement learning algorithm to construct a dynamic risk assessment model, which can comprehensively consider the severity, development trend and impact range of the fault, realize accurate assessment and dynamic adjustment of the risk level, and make the early warning information more accurate and reliable. The introduction of the distributed edge computing architecture makes full use of the computing resources of the ground station and the satellite end, significantly improves the real-time processing capability and scalability of the system, and realizes the local analysis and rapid response of the measurement and control data.

[0058] The present invention forms a complete set of technical solutions. While improving the accuracy of anomaly detection, it also realizes the intelligentization of fault diagnosis and the automation of early warning response, providing strong technical support for ensuring the safe and stable operation of satellite measurement and control systems. The solution has strong practicality and promotion value, and can be promoted and applied to related fields such as on-orbit measurement and control of spacecraft and aircraft health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a process for detecting satellite measurement and control data anomalies and for intelligent diagnosis and early warning according to an embodiment of the present invention;

[0060] Figure 2 The present invention is a schematic diagram of the structure of the satellite measurement and control data anomaly detection and intelligent diagnosis and early warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a method for detecting abnormal satellite measurement and control data and for intelligent diagnosis and early warning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S1. Acquire multi-source heterogeneous data in the satellite measurement and control system, wherein the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization and time alignment; construct a multi-level spatiotemporal fusion network model, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; based on the extracted high-dimensional spatiotemporal features, design a dynamic adaptive anomaly detection algorithm to perform anomaly detection on the satellite measurement and control data;

[0065] S2. Based on the anomaly detection results, use the pre-built multi-level satellite system knowledge graph to perform intelligent diagnosis; based on the intelligent diagnosis results, use the deep reinforcement learning algorithm to build a dynamic risk assessment model and generate multi-level warning information; design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to the ground station and satellite end;

[0066] S3. Build a human-machine collaborative decision-making system to present the generated multi-level warning information through a visual interface, and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on the operator's feedback information and historical data.

[0067] The present invention provides a satellite measurement and control data anomaly detection and intelligent diagnosis and early warning method, and the specific implementation method is as follows:

[0068] First, multi-source heterogeneous data is obtained from the satellite measurement and control system, including satellite telemetry data, orbit data, attitude data and ground station equipment status data. Satellite telemetry data contains parameters such as temperature, voltage, and current; orbit data contains information such as position and speed; attitude data contains attitude angle, angular velocity, etc.; ground station equipment status data includes antenna pointing, transmission power, etc.

[0069] Preprocess the acquired multi-source heterogeneous data. First, clean the data to remove outliers and missing values, such as removing values ​​in the temperature data that are outside the range of -100℃ to 100℃. Then normalize the different types of data and map the values ​​to the 0-1 interval. Finally, time alignment is performed to unify data with different sampling frequencies into the same time series, such as aligning 1Hz telemetry data and 10Hz attitude data into a 1s interval sequence.

[0070] Construct a multi-level spatiotemporal fusion network model. The model contains three key layers: a bidirectional gated recurrent unit layer, which is used to process time series data and extract time dimension features; a graph convolutional network layer, which is used to process spatial data and extract spatial dimension features; and an attention mechanism layer, which is used to fuse spatiotemporal features. Specifically, the bidirectional gated recurrent unit layer contains 128 hidden units; the graph convolutional network layer contains 3 graph convolutional layers, each with 64 convolution kernels; and the attention mechanism layer uses a multi-head self-attention mechanism with 8 heads.

[0071] The preprocessed multi-source heterogeneous data are input into the constructed multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features. Time series data such as telemetry parameter sequences are input into the bidirectional gated recurrent unit layer, and spatial data such as the topological relationship between satellite components are input into the graph convolutional network layer. Finally, a 256-dimensional high-dimensional spatiotemporal feature vector is obtained through the fusion of the attention mechanism layer.

[0072] Based on the extracted high-dimensional spatiotemporal features, a dynamic adaptive anomaly detection algorithm is designed. The algorithm uses a density-based local anomaly factor method to detect anomalies by calculating the local density ratio of a sample point and its k nearest neighbors. The key parameter k of the algorithm is dynamically adjusted over time, with an initial value set to 20 and gradually optimized according to the accuracy of the detection results. When an anomaly is detected, the ID of the abnormal sample and its anomaly score are output.

[0073] Intelligent diagnosis is performed using a pre-built multi-level satellite system knowledge graph. The knowledge graph contains entities and their relationships at multiple levels, such as satellite systems, subsystems, equipment, and parameters. Based on the anomaly detection results, reasoning is performed in the knowledge graph to locate possible fault causes. For example, if an abnormal current is detected on a solar panel, the knowledge graph can be used to trace possible fault causes such as solar panel damage and abnormal attitude.

[0074] Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm to generate multi-level warning information. The deep Q network algorithm is used, with the satellite status as input and the warning level as output, to continuously optimize the strategy network through interaction with the environment. The warning levels are divided into four levels: prompt, warning, serious, and emergency, corresponding to different disposal measures.

[0075] Design a distributed edge computing architecture and deploy some data processing and anomaly detection tasks to ground stations and satellites. Deploy lightweight anomaly detection models on satellites to monitor key parameters in real time; deploy preprocessing and feature extraction modules on ground stations to reduce the burden on central stations; and deploy complex diagnosis and early warning models at central stations. Each node uses federated learning to coordinate training and protect data privacy.

[0076] Build a human-machine collaborative decision-making system to present the generated multi-level warning information through a visual interface. The interface adopts the form of a dashboard to intuitively display the status of each satellite system and warning information. According to the warning level, the corresponding emergency plan is automatically triggered, such as adjusting the working mode, switching backup components, etc. The operator can adjust and confirm the plan according to the actual situation.

[0077] Design an incremental learning mechanism to continuously optimize each model based on operator feedback and historical data. Collect feedback data regularly, such as false positives and false negatives, and use transfer learning methods to update the parameters of the multi-level spatiotemporal fusion network model in combination with newly added historical data. At the same time, dynamically adjust the threshold of the anomaly detection algorithm based on the detection effect, optimize the reward function of the risk assessment model, and achieve continuous evolution of the model.

[0078] Through the above steps, this method realizes intelligent anomaly detection, diagnosis and early warning of satellite measurement and control data, and improves the safety and reliability of the satellite measurement and control system. This method can be widely used in the on-orbit measurement and control and health management of various satellites.

[0079] In an optional embodiment,

[0080] A multi-level spatiotemporal fusion network model is constructed, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data, and an attention mechanism layer for feature fusion; the pre-processed multi-source heterogeneous data is input into the multi-level spatiotemporal fusion network model, and the steps of extracting high-dimensional spatiotemporal features include:

[0081] The bidirectional gated recurrent unit layer processes past and future context information simultaneously through forward and backward gated recurrent units to capture the long-term dependencies of satellite states. The graph convolutional network layer represents the connection relationship between the components of the satellite system based on a predefined adjacency matrix and processes non-Euclidean spatial data through graph convolution operations. The attention mechanism layer uses a multi-head self-attention mechanism to learn the importance weights between different features.

[0082] The pre-processed multi-source heterogeneous satellite measurement and control data are input into the multi-level spatiotemporal fusion network model, where the time series data is input into the bidirectional gated recurrent unit layer, and the spatial data is input into the graph convolutional network layer;

[0083] Use the bidirectional gated recurrent unit layer to process the time series data to obtain the time series features; use the graph convolutional network layer to process the spatial data to obtain the spatial features;

[0084] The obtained temporal features and spatial features are concatenated, including: concatenating the temporal feature matrix and the spatial feature matrix on the feature dimension to form a comprehensive feature matrix, in which each row represents a time step or a spatial node, and each column represents a feature dimension; the concatenated temporal features and spatial features are input into the attention mechanism layer, and the fused features are obtained through multi-head self-attention calculation.

[0085] In this implementation, a multi-level spatiotemporal fusion network model is constructed to process multi-source heterogeneous data of the satellite measurement and control system. The model contains three key components: a bidirectional gated recurrent unit layer, a graph convolutional network layer, and an attention mechanism layer, which are used to process time series data, spatial data, and feature fusion, respectively.

[0086] First, the original satellite tracking data is preprocessed. This includes steps such as data cleaning, normalization, and time alignment to ensure data quality and make it suitable for input into the neural network model. For example, for temperature data, min-max normalization can be used to map it to the range of 0-1; for discrete state data, one-hot encoding can be performed to convert it into numerical form.

[0087] Next, the preprocessed data is divided into two parts: time series data and spatial data. Time series data includes telemetry data of satellite components that change over time, such as temperature, voltage, current, etc.; spatial data describes the topological relationship between the components of the satellite system.

[0088] For time series data, it is input to the bidirectional gated recurrent unit layer for processing. This layer contains two gated recurrent units, forward and backward, which process the sequence data from left to right and from right to left respectively. This bidirectional structure can capture past and future contextual information at the same time and effectively extract long-term dependencies. For example, for a temperature sequence of length 100, the forward unit starts processing from the first time step, while the backward unit starts processing from the 100th time step. Finally, the outputs of the two directions are fused to obtain a more comprehensive time series feature representation.

[0089] For spatial data, it is input into the graph convolutional network layer for processing. First, an adjacency matrix needs to be defined to represent the connection relationship between the components of the satellite system. For example, for a satellite system containing 10 nodes, a 10x10 matrix can be used to represent the connection between nodes, where the matrix element 1 represents connection and 0 represents disconnection. The graph convolution operation aggregates and updates the features of each node based on this adjacency matrix, thereby effectively processing spatial data with non-Euclidean structure. Through multi-layer graph convolution, the receptive field can be gradually expanded to capture a wider range of spatial dependencies.

[0090] The bidirectional gated recurrent unit layer and the graph convolutional network layer output temporal features and spatial features respectively. Next, these two features need to be fused. First, the temporal feature matrix and the spatial feature matrix are concatenated in the feature dimension to form a comprehensive feature matrix. Assuming that the temporal feature dimension is 100x64 and the spatial feature dimension is 10x32, the dimension of the concatenated comprehensive feature matrix is ​​110x96. Each row represents a time step or spatial node, and each column represents a feature dimension.

[0091] The concatenated comprehensive features are input to the attention mechanism layer. This layer uses a multi-head self-attention mechanism to learn the importance weights between different features. Specifically, the input features are first linearly projected into the query, key, and value spaces. Then the similarity between the query and the key is calculated to obtain the attention score, and the attention score is multiplied by the value to obtain the weighted feature. Through parallel calculation of multiple attention heads, the relationship between features can be learned from different representation subspaces. Finally, the outputs of multiple heads are concatenated and linearly transformed to obtain the final fused features.

[0092] Through the above steps, the multi-level spatiotemporal fusion network model can effectively extract and fuse high-dimensional spatiotemporal features in satellite measurement and control data. This model structure that integrates time series, space and attention mechanisms can make full use of the information in multi-source heterogeneous data and provide powerful feature representation for subsequent fault diagnosis and predictive maintenance tasks.

[0093] In an optional embodiment,

[0094] The method further comprises:

[0095] Constructing a multi-layer perceptron structure, wherein the multi-layer perceptron structure includes three fully connected layers, each of which is followed by a nonlinear activation function;

[0096] The fused features are input into the first fully connected layer of the multi-layer perceptron structure, the input fused features are linearly transformed, and nonlinearly mapped through a nonlinear activation function to obtain the first layer output features; the first layer output features are input into the second fully connected layer, and a second linear transformation and nonlinear mapping are performed to obtain the second layer output features; a Dropout layer is added between the second layer output features and the third fully connected layer; the second layer output features processed by Dropout are input into the third fully connected layer, and a final linear transformation is performed to obtain high-dimensional spatiotemporal features; wherein the nonlinear activation function adopts a rectified linear unit function.

[0097] In this embodiment, a multi-layer perceptron structure is constructed to further process the fusion features. The multi-layer perceptron structure includes three fully connected layers, each of which is followed by a nonlinear activation function. The specific implementation process is as follows:

[0098] First, a three-layer fully connected network structure is constructed. The input dimension of the first fully connected layer is the same as the dimension of the fusion feature, and the output dimension is set to 1024. The input dimension of the second fully connected layer is 1024, and the output dimension is set to 512. The input dimension of the third fully connected layer is 512, and the output dimension is set to 256. Each fully connected layer is followed by a ReLU activation function as a nonlinear transformation.

[0099] Next, the fused features are input to the first fully connected layer. Assuming that the dimension of the fused features is 2048, the dimension of the weight matrix of the first fully connected layer is 2048×1024, and the dimension of the bias vector is 1024. The fused features are multiplied by the weight matrix, and then the bias vector is added to obtain the result after linear transformation. The result is then nonlinearly mapped through the ReLU function to obtain the output features of the first layer. The function of the ReLU function is to set negative values ​​to 0 and keep positive values ​​unchanged, which can increase the nonlinear expression ability of the network.

[0100] Then, the output features of the first layer are input to the second fully connected layer. The weight matrix dimension of the second fully connected layer is 1024×512, and the bias vector dimension is 512. Similarly, linear transformation and ReLU nonlinear mapping are performed to obtain the output features of the second layer.

[0101] Add a Dropout layer between the second and third fully connected layers. The Dropout layer randomly sets a certain proportion (such as 0.5) of the neuron outputs to 0, which can reduce overfitting and improve the generalization ability of the model. Assuming that the dimension of the second layer output feature is 512, after the Dropout process, the outputs of about 256 neurons are set to 0.

[0102] Finally, the features processed by Dropout are input to the third fully connected layer. The weight matrix dimension of the third fully connected layer is 512×256, and the bias vector dimension is 256. After the last linear transformation, a 256-dimensional high-dimensional spatiotemporal feature is obtained. This high-dimensional spatiotemporal feature fuses the information of the original fused features and extracts a more abstract feature representation through multiple layers of nonlinear transformation.

[0103] Through this multi-layer perceptron structure, deep feature extraction and nonlinear mapping can be performed on the fusion features to obtain more expressive high-dimensional spatiotemporal features. This high-dimensional spatiotemporal feature can be used for subsequent classification or regression tasks, such as behavior recognition, trajectory prediction, etc. In the whole process, the use of ReLU activation function enhances the nonlinear fitting ability of the network, and the introduction of Dropout helps to improve the generalization performance of the model.

[0104] In an optional embodiment,

[0105] Based on the extracted high-dimensional spatiotemporal features, a dynamic adaptive anomaly detection algorithm is designed. The steps of anomaly detection on satellite measurement and control data include:

[0106] A dynamic adaptive anomaly detection algorithm is designed by combining a mixed Gaussian model with a dynamic threshold; the mixed Gaussian model models the probability distribution of high-dimensional spatiotemporal features, and the parameters of the mixed Gaussian model are estimated by an expectation maximization algorithm. The mixed Gaussian model includes multiple Gaussian components, each of which has a corresponding mixing coefficient, a mean vector, and a covariance matrix;

[0107] Calculating the negative log-likelihood score of the high-dimensional spatiotemporal features using the mixed Gaussian model, setting a time window, calculating the mean and standard deviation of the negative log-likelihood score within the time window, and constructing a dynamic threshold based on the mean and standard deviation;

[0108] Perform an exponentially weighted moving average on the negative log-likelihood scores of multiple consecutive time steps to obtain a smoothed score sequence; calculate the Mahalanobis distance of the high-dimensional spatiotemporal features, where the Mahalanobis distance is calculated based on the mean vector and the covariance matrix; perform a weighted combination of the smoothed score sequence and the Mahalanobis distance to construct a composite anomaly metric; compare the composite anomaly metric with a dynamic threshold, and when the composite anomaly metric exceeds the dynamic threshold, mark the current time step as an anomaly point; analyze the contribution of the high-dimensional spatiotemporal features marked as anomalies in each feature dimension, and identify the feature dimensions that cause the anomaly.

[0109] In a specific implementation, the process of performing anomaly detection on satellite measurement and control data based on a dynamic adaptive anomaly detection algorithm combined with a mixed Gaussian model and a dynamic threshold is as follows:

[0110] When designing a dynamic adaptive anomaly detection algorithm, the extracted high-dimensional spatiotemporal features are first modeled using a mixed Gaussian model. The mixed Gaussian model contains multiple Gaussian components, each of which has its corresponding mixing coefficient, mean vector, and covariance matrix. These parameters are estimated using the expectation maximization iterative algorithm, which randomly initializes the parameters and then alternates between the expectation step and the maximization step until convergence. Taking ten features such as satellite temperature, voltage, and current as examples, three Gaussian components can be set, and their mixing coefficients are 0.4, 0.35, and 0.25, respectively.

[0111] For each high-dimensional feature vector at each time step, the trained mixed Gaussian model is used to calculate its negative log-likelihood score. A 60-second sliding time window is set, and the mean and standard deviation of the score sequence are calculated within the window to construct a dynamic threshold. The threshold calculation considers the mean plus or minus the multiple of the standard deviation. The multiple factor can be adjusted dynamically according to actual needs. For example, when the value is 3, it means that samples outside the range of plus or minus three times the standard deviation of the mean are considered potential anomalies.

[0112] To reduce the impact of noise, the negative log-likelihood scores of 10 consecutive time steps are exponentially weighted moving averaged, with the decay factor set to 0.7, and newer scores have greater weights. At the same time, the Mahalanobis distance from the eigenvector to the center of each Gaussian component is calculated, taking the covariance structure into account. The smoothed score sequence and the Mahalanobis distance are weighted and combined with weights of 0.6 and 0.4 to obtain a composite anomaly measure.

[0113] Compare the composite anomaly metric with the dynamic threshold. When the metric exceeds the threshold, mark the time step as an abnormal point. Taking the temperature sensor data as an example, if the composite anomaly metric at a certain moment is 15.8, and the current dynamic threshold is 12.3, then the point is judged to be abnormal. Further analyze the feature contribution of the abnormal point, calculate the degree of deviation of each dimension of the feature from the normal mode, and find out the key feature dimensions that cause the abnormality. For example, it may be found that the deviation of the two features of temperature and current is the largest, reaching 2.8 and 2.5 times the standard deviation respectively.

[0114] Through the design of dynamic thresholds, the algorithm can adapt to changes in data distribution. When the system enters a new working mode, the threshold will be adjusted accordingly. Feature contribution analysis provides important clues for abnormal diagnosis and helps to locate the source of the fault. The advantages of this technical solution are: combining probability models and distance metrics to improve the accuracy of anomaly detection; using dynamic thresholds and smoothing to enhance the robustness of the algorithm; and providing interpretable abnormal diagnosis results through feature contribution analysis. Experiments show that this solution can effectively detect abnormal patterns in satellite measurement and control data and provide strong support for spacecraft health monitoring.

[0115] In an optional embodiment,

[0116] Based on the anomaly detection results, the steps for intelligent diagnosis using the pre-built multi-level satellite system knowledge graph include:

[0117] Construct a multi-level satellite system knowledge graph containing physical device nodes, functional module nodes, state parameter nodes, fault mode nodes and symptom feature nodes, and establish physical inclusion relationships, functional dependency relationships, parameter association relationships, fault propagation relationships and symptom attribution relationships between nodes;

[0118] The multi-level satellite system knowledge graph is modeled by using a graph attention network, the attention weights between nodes are calculated through a multi-head attention mechanism, and the feature representation of the nodes is updated according to the attention weights; the anomaly detection results are input into the graph attention network, the corresponding symptom feature nodes are activated, and the activation strength of the symptom nodes is obtained; based on the activation strength of the symptom nodes, the fault propagation process is simulated in the multi-level satellite system knowledge graph, and the activation state of each fault node is calculated; according to the activation state of the fault node, the root cause of the fault is determined by reverse reasoning, and the association strength of the root cause node is obtained; according to the association strength of the root cause node, the path score connecting the symptom node and the fault node is calculated, and the path above the importance threshold is extracted to form a diagnostic subgraph;

[0119] Based on the diagnosis subgraph, evidence sources are extracted from three dimensions: physical equipment, functional modules and state parameters, and probability weights are assigned to each evidence source according to the importance of each dimension. The Dempster combination rule is used to fuse the evidence sources to obtain the final diagnosis result.

[0120] According to the anomaly detection results, the specific implementation method of intelligent diagnosis using the pre-built multi-level satellite system knowledge graph is as follows:

[0121] First, a multi-level satellite system knowledge graph is constructed. The knowledge graph contains five types of nodes: physical device nodes, functional module nodes, state parameter nodes, fault mode nodes, and symptom feature nodes. Physical device nodes represent specific hardware devices in the satellite system, such as solar panels, attitude control devices, etc. Functional module nodes represent functional units of the satellite system, such as power supply systems, attitude control systems, etc. State parameter nodes represent key indicators during the operation of the satellite system, such as voltage, temperature, angular velocity, etc. Fault mode nodes represent possible fault types, such as short circuit, overheating, deflection, etc. Symptom feature nodes represent the external manifestations of faults, such as voltage fluctuations, abnormal temperatures, unstable attitude, etc.

[0122] Five kinds of relationships are established between nodes: the physical inclusion relationship represents the hierarchical structure between devices, such as the solar panels are included in the power system; the functional dependency relationship represents the functional association between modules, such as the attitude control system depends on the power system; the parameter association relationship represents the mutual influence between parameters, such as the association between voltage and temperature; the fault propagation relationship represents the diffusion path of the fault, such as short circuit may cause overheating; the symptom attribution relationship represents the correspondence between symptoms and faults, such as voltage fluctuation is a symptom of short circuit.

[0123] Through the above method, a multi-level satellite system knowledge graph containing 3,000 nodes and 10,000 edges can be constructed to comprehensively characterize the composition structure, functional relationships and failure modes of the satellite system.

[0124] Next, the multi-level satellite system knowledge graph is modeled using a graph attention network. First, each node is initialized as a 64-dimensional feature vector, and the vector elements are between [-1, 1]. Then the attention weights between nodes are calculated through the 8-head attention mechanism. For any two connected nodes, their query vectors, key vectors, and value vectors are calculated respectively. The dot product of the query vector and the key vector is used to obtain the attention score, which is then normalized by softmax to obtain the attention weight. The value vector is weighted and summed according to the attention weight to obtain the updated node features. Repeat this process 3 times to obtain the final feature representation of the node.

[0125] The anomaly detection results are input into the graph attention network to activate the corresponding symptom feature nodes. For example, when voltage fluctuation is detected, the activation position in the feature vector of the voltage fluctuation node is set to 1, and the other positions are set to 0. Through the propagation of the graph attention network, the nodes related to voltage fluctuation will also be activated, and the activation strength will decay according to the attention weight. Finally, the activation strength of all symptom nodes is obtained, with a value between [0,1].

[0126] Based on the activation strength of symptom nodes, the fault propagation process is simulated in the multi-level satellite system knowledge graph. Starting from the activated symptom node, the fault propagation is propagated upward along the symptom attribution relationship and the fault propagation relationship, and the activation state of each fault node is calculated. During the propagation process, the activation strength of the child node will be attenuated and transferred to the parent node, and the attenuation coefficient is set to 0.8. Repeat the propagation until the activation state of all nodes converges.

[0127] According to the activation state of the fault node, the root cause of the fault is determined by reverse reasoning. Starting from the fault node with the highest activation strength, trace upward along the fault propagation relationship until a node with a sudden change in activation strength is found, which is the possible root cause node. Calculate the association strength between the root cause node and the initial symptom node, which is equal to the product of the activation strengths of all edges on the path.

[0128] According to the correlation strength of the root cause node, the path score connecting the symptom node and the fault node is calculated. The path score is equal to the sum of the activation strengths of all nodes on the path. The importance threshold is set to the average value of the path score, and the paths above the threshold are extracted to form a diagnostic subgraph. The diagnostic subgraph contains the most relevant symptom, fault, and root cause information.

[0129] Based on the diagnostic subgraph, evidence sources are extracted from three dimensions: physical devices, functional modules, and state parameters. The three device nodes with the highest activation intensity are selected from the physical device dimension, the three module nodes with the highest activation intensity are selected from the functional module dimension, and the five parameter nodes with the highest activation intensity are selected from the state parameter dimension. According to the importance of each dimension, probability weights of 0.3, 0.3, and 0.4 are assigned to physical devices, functional modules, and state parameters, respectively.

[0130] Finally, the Dempster combination rule (DS evidence theory combination rule) is used to fuse the evidence sources. First, the basic probability distribution of each evidence source is calculated, and then the probability distribution is updated by combining them two by two, and finally the fused probability distribution is obtained. The fault type with the highest probability is the final diagnosis result. For example, the diagnosis result shows that the probability of short circuit fault is 0.8, the probability of overheating fault is 0.15, and the probability of other faults is 0.05.

[0131] Through the above steps, based on the anomaly detection results and the multi-level satellite system knowledge graph, the intelligent diagnosis of satellite system faults can be realized, and decision support can be provided for fault location and maintenance. This method makes comprehensive use of domain knowledge and data-driven methods, and has strong interpretability and accuracy.

[0132] In an optional embodiment,

[0133] Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm. The steps to generate multi-level warning information include:

[0134] Acquire real-time diagnostic results and system operation data, convert the real-time diagnostic results into binary feature vectors through one-hot encoding, fuse the system operation data with the binary feature vectors after standardization to obtain fused features; establish an action space including risk level assessment and warning level selection;

[0135] Construct a dual deep Q network structure, including a main network and a target network. Both the main network and the target network are composed of a multi-channel input layer, a convolutional feature extraction layer, a fully connected fusion layer, and an action value output layer.

[0136] Calculate the loss function including the temporal difference error, use the stochastic gradient descent method to update the main network parameters, and periodically synchronize the main network parameters to the target network through the soft update strategy;

[0137] Based on the fusion features, the Q value of each action is calculated through the main network, and the evaluation action is selected to determine the risk level and warning level, generating multi-level warning information including risk determination basis, potential hazard analysis and disposal suggestions.

[0138] The method of building a dynamic risk assessment model based on intelligent diagnosis results and generating multi-level warning information is implemented as follows:

[0139] First, obtain the real-time diagnostic results and system operation data. Real-time diagnostic results usually include the results of multiple diagnostic items, such as "bearing abnormality" and "gear wear". These diagnostic results are converted into binary feature vectors through one-hot encoding. Specifically, assuming that there are 5 diagnostic items and a certain diagnostic result is "bearing abnormality", the corresponding binary feature vector is [1,0,0,0,0]. System operation data includes multiple parameters such as temperature, vibration, and pressure. These data are standardized so that their numerical range is unified between 0 and 1. Then the standardized system operation data is spliced ​​with the binary feature vector to obtain a fused feature vector.

[0140] Next, we build an action space, which includes two dimensions: risk level assessment and warning level selection. Risk levels can be divided into three levels: low, medium, and high, and warning levels can be divided into three levels: general warning, important warning, and emergency warning. Therefore, there are 9 combinations in the action space.

[0141] Construct a dual deep Q network structure, including the main network and the target network. The two networks have the same structure, consisting of a multi-channel input layer, a convolutional feature extraction layer, a fully connected fusion layer, and an action value output layer. The multi-channel input layer receives the fused feature vector, the convolutional feature extraction layer uses multiple convolution kernels to extract features, the fully connected fusion layer fuses the extracted features, and finally the output layer outputs the Q values ​​of 9 actions.

[0142] During the training process, the Q value of each action is first calculated through the main network based on the current state (fusion features), and the action with the largest Q value is selected as the evaluation action. After executing the action, the reward and the next state are obtained, and the maximum Q value of each action in the next state is calculated through the target network. The loss function containing the temporal difference error is calculated, and the parameters of the main network are updated using the stochastic gradient descent method. At regular intervals, the parameters of the main network are periodically synchronized to the target network through the soft update strategy, and the soft update coefficient is set to 0.01.

[0143] After the model training is completed, in actual application, based on the fusion features obtained in real time, the Q value of each action is calculated through the main network, and the action with the largest Q value is selected as the evaluation result to determine the risk level and warning level. For example, if the evaluation result is high risk and emergency warning, the corresponding warning information will be generated, including the basis for risk determination (such as bearing abnormality, high temperature), potential hazard analysis (such as possible equipment shutdown) and disposal suggestions (such as immediate shutdown for maintenance).

[0144] Through the above steps, a dynamic risk assessment model is constructed based on the intelligent diagnosis results, which can evaluate the equipment status in real time and generate multi-level warning information, providing strong support for equipment maintenance decisions. This method combines deep learning and reinforcement learning technology and has strong adaptive ability and generalization performance.

[0145] In an optional embodiment,

[0146] The steps to design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to ground stations and satellites include:

[0147] A ground station distributed computing framework including a data preprocessing layer, a model training layer, and an inference service layer is constructed. The data preprocessing layer distinguishes duplicate data through a Bloom filter and performs distributed processing of data cleaning and feature engineering with an incremental storage mechanism to generate a standardized training data set; the model training layer uses a parameter server architecture and a ring full reduction algorithm to perform model parallel training based on the training data set, and migrates the trained complex model knowledge to a lightweight model through a model distillation method, which is deployed to the inference service layer for online inference; the inference service layer monitors and counts the computing load indicator information of each computing node in real time, and the computing load indicator information includes CPU usage, memory occupancy, and task queue length;

[0148] Establish an edge node task scheduling mechanism, receive the computing load indicator information of the inference service layer, and allocate new tasks according to the task complexity, network quality and node computing power. When the task complexity is lower than the preset threshold, it is allocated to the satellite end; when the network quality is lower than the preset threshold, it is allocated to the nearest ground station.

[0149] Perform data compression and incremental update processing on the assigned tasks, compress the transmitted data according to the adaptive compression ratio, and obtain the status of each node through the lightweight heartbeat protocol; generate checkpoints and snapshot backups during task execution based on the node status, and read the task status from the most recent checkpoint data when a node failure is detected, and migrate the task to other idle nodes;

[0150] The task data is stored in a multi-copy mechanism, and the task execution results are transmitted to the data preprocessing layer in an incremental update manner for updating the training data set.

[0151] For example, when designing a distributed edge computing architecture, first build a ground station distributed computing framework that includes a data preprocessing layer, a model training layer, and an inference service layer. The data preprocessing layer uses a Bloom filter to perform duplicate detection on the input data and stores the new data in an incremental manner. Then, data cleaning is performed to remove outliers and missing values, and feature engineering is performed, such as one-hot encoding and standardization, to finally generate a standardized training data set.

[0152] The model training layer uses a parameter server architecture for distributed training based on the preprocessed training data set. Specifically, the model parameters are stored on the parameter server, and multiple worker nodes calculate gradients and update parameters in parallel. The ring full-reduce algorithm is used to synchronize gradients between worker nodes to improve communication efficiency. For example, for a cluster with 8 worker nodes, it can be organized into a ring topology, and each worker node only needs to communicate with two adjacent nodes to complete gradient synchronization. After training, the knowledge of the complex model is transferred to the lightweight model through the knowledge distillation method. Specifically, the teacher-student architecture can be used, with the complex model as the teacher model and the lightweight model as the student model. The student model is trained by minimizing the relative entropy of the outputs of the two.

[0153] The inference service layer deploys trained lightweight models and monitors the load of each computing node in real time. By regularly collecting indicators such as CPU usage, memory occupancy, and task queue length, a load statistics report is constructed. For example, data can be collected every 5 seconds, and the average value of the last minute is calculated as the current load indicator.

[0154] On this basis, an edge node task scheduling mechanism is established. The scheduler receives the load indicator information reported by the inference service layer, and allocates tasks based on factors such as the complexity of the newly added tasks and the current network quality. Specifically, the task complexity is first evaluated, which can be quantified by estimating the amount of computation and memory usage. If the complexity is lower than the preset threshold (such as the estimated execution time <100ms), the task is assigned to the satellite for execution. Secondly, the current network quality is detected. If it is lower than the preset threshold (such as RTT>200ms), the task is assigned to the nearest ground station. For other cases, the appropriate execution node is selected according to the load balancing principle.

[0155] Perform data compression and incremental update processing on the assigned tasks. Adopt an adaptive compression algorithm to dynamically adjust the compression ratio according to the network bandwidth. For example, when the bandwidth is sufficient, use lightweight compression such as the Snappy algorithm, and when the bandwidth is limited, use the LZMA2 algorithm with a higher compression ratio. Monitor the status of each node through a lightweight heartbeat protocol. The heartbeat packet contains information such as node ID, timestamp, and load indicators. Based on the heartbeat feedback, checkpoints are generated regularly during the task execution process to record the current execution status and intermediate results. At the same time, task snapshots are created to save the complete execution environment. When a node failure is detected, the task status is restored from the most recent checkpoint, and the task is migrated to other idle nodes for continued execution.

[0156] A multi-copy mechanism is used to store task data to improve reliability. For example, three copies of important data are maintained and distributed on different nodes. After the task is completed, the result is transmitted to the data preprocessing layer in an incremental update manner. Specifically, only the difference with the previous version is transmitted to reduce the amount of data transmission. After receiving the incremental update, the data preprocessing layer merges it with the original data, updates the training data set, and provides the latest data support for subsequent model training.

[0157] Through the above steps, an efficient and reliable distributed edge computing architecture is realized, which can flexibly schedule tasks, balance loads, ensure fault tolerance, and continuously optimize model performance. This architecture is suitable for large-scale data processing and real-time anomaly detection tasks in scenarios such as satellite remote sensing.

[0158] Figure 2 FIG. 1 is a schematic diagram of the structure of a satellite measurement and control data anomaly detection and intelligent diagnosis and early warning system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0159] The first unit is used to obtain multi-source heterogeneous data in the satellite measurement and control system, and the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization and time alignment; construct a multi-level spatiotemporal fusion network model, and the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; based on the extracted high-dimensional spatiotemporal features, design a dynamic adaptive anomaly detection algorithm to perform anomaly detection on satellite measurement and control data;

[0160] The second unit is used to perform intelligent diagnosis based on the anomaly detection results using the pre-built multi-level satellite system knowledge graph; based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm to generate multi-level warning information; a distributed edge computing architecture is designed to deploy some data processing and anomaly detection tasks to ground stations and satellite terminals;

[0161] The third unit is used to build a human-machine collaborative decision-making system, which will present the generated multi-level warning information through a visual interface and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on the operator's feedback information and historical data.

[0162] According to a third aspect of the embodiments of the present invention,

[0163] An electronic device is provided, comprising:

[0164] processor;

[0165] a memory for storing processor-executable instructions;

[0166] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0167] According to a fourth aspect of the embodiments of the present invention,

[0168] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0169] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Satellite measurement and control data anomaly detection and intelligent diagnosis and early warning method, characterized in that: include: Acquire multi-source heterogeneous data in a satellite measurement and control system, wherein the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data, and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization, and time alignment; construct a multi-level spatiotemporal fusion network model, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data, and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; design a dynamic adaptive anomaly detection algorithm based on the extracted high-dimensional spatiotemporal features to perform anomaly detection on satellite measurement and control data; Based on the anomaly detection results, intelligent diagnosis is performed using the pre-built multi-level satellite system knowledge graph; Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using deep reinforcement learning algorithms to generate multi-level warning information; Design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to ground stations and satellites; Build a human-machine collaborative decision-making system to present the generated multi-level warning information through a visual interface, and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on operator feedback information and historical data.

2. The method according to claim 1, characterized in that Constructing a multi-level spatiotemporal fusion network model, wherein the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data, and an attention mechanism layer for feature fusion; The preprocessed multi-source heterogeneous data is input into the multi-level spatiotemporal fusion network model, and the steps of extracting high-dimensional spatiotemporal features include: The bidirectional gated recurrent unit layer processes past and future context information simultaneously through forward and backward gated recurrent units to capture the long-term dependencies of satellite states. The graph convolutional network layer represents the connection relationship between the components of the satellite system based on a predefined adjacency matrix and processes non-Euclidean spatial data through graph convolution operations. The attention mechanism layer uses a multi-head self-attention mechanism to learn the importance weights between different features. The pre-processed multi-source heterogeneous satellite measurement and control data are input into the multi-level spatiotemporal fusion network model, where the time series data is input into the bidirectional gated recurrent unit layer, and the spatial data is input into the graph convolutional network layer; Use the bidirectional gated recurrent unit layer to process the time series data to obtain the time series features; use the graph convolutional network layer to process the spatial data to obtain the spatial features; The obtained temporal features and spatial features are concatenated, including: concatenating the temporal feature matrix and the spatial feature matrix on the feature dimension to form a comprehensive feature matrix, in which each row represents a time step or a spatial node, and each column represents a feature dimension; the concatenated temporal features and spatial features are input into the attention mechanism layer, and the fused features are obtained through multi-head self-attention calculation.

3. The method according to claim 2, characterized in that The method further comprises: Constructing a multi-layer perceptron structure, wherein the multi-layer perceptron structure includes three fully connected layers, each of which is followed by a nonlinear activation function; The fused features are input into the first fully connected layer of the multi-layer perceptron structure, the input fused features are linearly transformed, and nonlinearly mapped through a nonlinear activation function to obtain the first layer output features; the first layer output features are input into the second fully connected layer, and a second linear transformation and nonlinear mapping are performed to obtain the second layer output features; a Dropout layer is added between the second layer output features and the third fully connected layer; the second layer output features processed by Dropout are input into the third fully connected layer, and a final linear transformation is performed to obtain high-dimensional spatiotemporal features; wherein the nonlinear activation function adopts a rectified linear unit function.

4. The method according to claim 1, characterized in that Based on the extracted high-dimensional spatiotemporal features, a dynamic adaptive anomaly detection algorithm is designed. The steps of anomaly detection on satellite measurement and control data include: A dynamic adaptive anomaly detection algorithm is designed by combining a mixed Gaussian model with a dynamic threshold; the mixed Gaussian model models the probability distribution of high-dimensional spatiotemporal features, and the parameters of the mixed Gaussian model are estimated by an expectation maximization algorithm. The mixed Gaussian model includes multiple Gaussian components, each of which has a corresponding mixing coefficient, a mean vector, and a covariance matrix; Calculating the negative log-likelihood score of the high-dimensional spatiotemporal features using the mixed Gaussian model, setting a time window, calculating the mean and standard deviation of the negative log-likelihood score within the time window, and constructing a dynamic threshold based on the mean and standard deviation; Perform an exponentially weighted moving average on the negative log-likelihood scores of multiple consecutive time steps to obtain a smoothed score sequence; calculate the Mahalanobis distance of the high-dimensional spatiotemporal features, where the Mahalanobis distance is calculated based on the mean vector and the covariance matrix; perform a weighted combination of the smoothed score sequence and the Mahalanobis distance to construct a composite anomaly metric; compare the composite anomaly metric with a dynamic threshold, and when the composite anomaly metric exceeds the dynamic threshold, mark the current time step as an anomaly point; analyze the contribution of the high-dimensional spatiotemporal features marked as anomalies in each feature dimension, and identify the feature dimensions that cause the anomaly.

5. The method according to claim 1, characterized in that Based on the anomaly detection results, the steps for intelligent diagnosis using the pre-built multi-level satellite system knowledge graph include: Construct a multi-level satellite system knowledge graph containing physical device nodes, functional module nodes, state parameter nodes, fault mode nodes and symptom feature nodes, and establish physical inclusion relationships, functional dependency relationships, parameter association relationships, fault propagation relationships and symptom attribution relationships between nodes; The multi-level satellite system knowledge graph is modeled by using a graph attention network, the attention weights between nodes are calculated through a multi-head attention mechanism, and the feature representation of the nodes is updated according to the attention weights; the anomaly detection results are input into the graph attention network, the corresponding symptom feature nodes are activated, and the activation strength of the symptom nodes is obtained; based on the activation strength of the symptom nodes, the fault propagation process is simulated in the multi-level satellite system knowledge graph, and the activation state of each fault node is calculated; according to the activation state of the fault node, the root cause of the fault is determined by reverse reasoning, and the association strength of the root cause node is obtained; according to the association strength of the root cause node, the path score connecting the symptom node and the fault node is calculated, and the path above the importance threshold is extracted to form a diagnostic subgraph; Based on the diagnosis subgraph, evidence sources are extracted from three dimensions: physical equipment, functional modules and state parameters, and probability weights are assigned to each evidence source according to the importance of each dimension. The Dempster combination rule is used to fuse the evidence sources to obtain the final diagnosis result.

6. The method according to claim 1, characterized in that Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using a deep reinforcement learning algorithm. The steps to generate multi-level warning information include: Acquire real-time diagnostic results and system operation data, convert the real-time diagnostic results into binary feature vectors through one-hot encoding, fuse the system operation data with the binary feature vectors after standardization to obtain fused features; establish an action space including risk level assessment and warning level selection; Construct a dual deep Q network structure, including a main network and a target network. Both the main network and the target network are composed of a multi-channel input layer, a convolutional feature extraction layer, a fully connected fusion layer, and an action value output layer. Calculate the loss function including the temporal difference error, use the stochastic gradient descent method to update the main network parameters, and periodically synchronize the main network parameters to the target network through the soft update strategy; Based on the fusion features, the Q value of each action is calculated through the main network, and the evaluation action is selected to determine the risk level and warning level, generating multi-level warning information including risk determination basis, potential hazard analysis and disposal suggestions.

7. The method according to claim 1, characterized in that The steps to design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to ground stations and satellites include: A ground station distributed computing framework including a data preprocessing layer, a model training layer, and an inference service layer is constructed. The data preprocessing layer distinguishes duplicate data through a Bloom filter and performs distributed processing of data cleaning and feature engineering with an incremental storage mechanism to generate a standardized training data set; the model training layer uses a parameter server architecture and a ring full reduction algorithm to perform model parallel training based on the training data set, and migrates the trained complex model knowledge to a lightweight model through a model distillation method, which is deployed to the inference service layer for online inference; the inference service layer monitors and counts the computing load indicator information of each computing node in real time, and the computing load indicator information includes CPU usage, memory occupancy, and task queue length; Establish an edge node task scheduling mechanism, receive the computing load indicator information of the inference service layer, and allocate new tasks according to the task complexity, network quality and node computing power. When the task complexity is lower than the preset threshold, it is allocated to the satellite end; when the network quality is lower than the preset threshold, it is allocated to the nearest ground station. Perform data compression and incremental update processing on the assigned tasks, compress the transmitted data according to the adaptive compression ratio, and obtain the status of each node through the lightweight heartbeat protocol; generate checkpoints and snapshot backups during task execution based on the node status, and read the task status from the most recent checkpoint data when a node failure is detected, and migrate the task to other idle nodes; The task data is stored in a multi-copy mechanism, and the task execution results are transmitted to the data preprocessing layer in an incremental update manner for updating the training data set.

8. A satellite measurement and control data anomaly detection and intelligent diagnosis and early warning system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source heterogeneous data in the satellite measurement and control system, and the multi-source heterogeneous data includes satellite telemetry data, orbit data, attitude data and ground station equipment status data; pre-process the multi-source heterogeneous data, including data cleaning, normalization and time alignment; construct a multi-level spatiotemporal fusion network model, and the multi-level spatiotemporal fusion network model includes a bidirectional gated recurrent unit layer for processing time series data, a graph convolutional network layer for processing spatial data and an attention mechanism layer for feature fusion; input the pre-processed multi-source heterogeneous data into the multi-level spatiotemporal fusion network model to extract high-dimensional spatiotemporal features; based on the extracted high-dimensional spatiotemporal features, design a dynamic adaptive anomaly detection algorithm to perform anomaly detection on satellite measurement and control data; The second unit is used to perform intelligent diagnosis based on the anomaly detection results using a pre-built multi-level satellite system knowledge graph; Based on the intelligent diagnosis results, a dynamic risk assessment model is constructed using deep reinforcement learning algorithms to generate multi-level warning information; Design a distributed edge computing architecture to deploy some data processing and anomaly detection tasks to ground stations and satellites; The third unit is used to build a human-machine collaborative decision-making system, which will present the generated multi-level warning information through a visual interface and automatically trigger the corresponding emergency plan according to the warning level; design an incremental learning mechanism to continuously optimize the multi-level spatiotemporal fusion network model, dynamic adaptive anomaly detection algorithm and dynamic risk assessment model based on the operator's feedback information and historical data.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images