Cloud edge collaboration-based low earth orbit satellite terminal fault prediction method and system and medium

Abstract

The application relates to a low-orbit satellite terminal fault prediction method and system based on cloud-edge cooperation and a medium, the low-orbit satellite terminal fault prediction method based on cloud-edge cooperation comprises the following steps: S1, converting multi-source heterogeneous original data generated by a low-orbit satellite terminal into a standardized time sequence feature matrix; S2, constructing a dynamic space-time graph neural network model as a teacher model for capturing complex space-time correlation characteristics among all network terminals; S3, converting the teacher model into a lightweight student model suitable for deployment of the low-orbit satellite terminal; S4, loading the quantized student model of the low-orbit satellite terminal to perform real-time online reasoning; S5, cloud-edge cooperative interaction, adopting an event-driven mechanism, and starting uplink transmission of the low-orbit satellite terminal only when the warning in S4 is triggered; and S6, converting a data-driven prediction result into a diagnosis report by using a domain knowledge graph. The application significantly improves the accuracy, real-time performance and interpretability of fault prediction, and provides strong support for operation and maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of satellite internet communication. In particular, it relates to a method, system, and medium for predicting faults in low-Earth orbit satellite terminals based on cloud-edge collaboration. Background Technology

[0002] Fault prediction for low-Earth orbit satellite internet terminals is a key technology for ensuring the stable operation of large-scale constellation networks.

[0003] The existing technical solutions are mainly divided into four aspects: (1) Alarm scheme based on rules and static thresholds: This scheme uses statistical methods such as time series analysis and regression analysis to mine trends and abnormal patterns in the data. It is suitable for linear or simple nonlinear relationships, but it is difficult to effectively capture complex coupling relationships and high-order fault modes between variables, and its ability to predict sudden faults is limited, making it difficult to provide accurate explanations of the root causes of faults. (2) Prediction scheme based on statistics and time series analysis: This scheme uses statistical methods such as time series analysis and regression analysis to mine trends and abnormal patterns in the data. It is suitable for linear or simple nonlinear relationships, but it is difficult to effectively capture complex coupling relationships and high-order fault modes between variables, and its ability to predict sudden faults is limited, making it difficult to provide accurate explanations of the root causes of faults. (3) Classification and regression scheme based on traditional machine learning: This scheme regards fault prediction as a supervised learning problem and trains classification or regression models through data after feature selection. Usually, this scheme focuses on modeling a single terminal and ignores the strong spatiotemporal correlation between terminals in the low-orbit satellite network, making it unable to effectively warn of group risks. Moreover, since the training process heavily relies on a large number of balanced labeled fault samples, the generalization ability is poor in small sample scenarios. (4) End-to-end prediction scheme based on deep learning: This scheme uses deep neural networks (such as LSTM, CNN) to automatically learn deep features in massive time series data, but requires a large amount of training data and high computing resources. Although it can handle multivariate time series problems of a single terminal, it lacks a mechanism for explicitly modeling the dynamic spatiotemporal relationship between terminals. Therefore, its prediction perspective is limited.

[0004] Current low-Earth orbit (LEO) satellite internet terminal fault prediction technologies suffer from the following major shortcomings. First, existing methods are largely limited to isolated analysis of device logs or single link indicators, failing to effectively integrate multi-source heterogeneous data, such as device status parameters, link quality fluctuations, terminal spatiotemporal trajectories, and electromagnetic interference. Since LEO satellite terminal faults are typically caused by the coupling of multiple factors, a single data perspective struggles to identify subtle pre-fault signs, resulting in insufficient prediction timeliness. Second, existing solutions usually model each terminal as an independent entity, neglecting the strong spatiotemporal correlations between terminals in the LEO satellite network. This isolated modeling approach cannot effectively utilize the status information of neighboring terminals to assist in fault prediction, making it difficult to achieve early warning of group faults. Third, deep learning-based fault prediction models typically have a large number of parameters and high computational complexity, requiring high-performance GPUs. However, LEO satellite terminals have limited computing resources and power consumption, making it difficult to directly run large-scale cloud models, resulting in poor real-time prediction capabilities or reliance on high-bandwidth backhaul of all raw data. Finally, most existing deep learning models only output fault probabilities, lacking in-depth analysis of the root causes of faults and supporting evidence chains. This makes it difficult for operations and maintenance personnel to trust the forecast results, thus hindering the development of targeted maintenance strategies and impacting the efficiency of translating forecasts into actual operations and maintenance actions. Addressing these shortcomings is a pressing issue that needs to be resolved. Summary of the Invention

[0005] This invention provides a method, system, and medium for predicting faults in low-orbit satellite terminals based on cloud-edge collaboration, in order to address the issues of low prediction accuracy and timeliness, as well as poor early warning capabilities for mass faults.

[0006] To achieve the above objectives, in a first aspect, the present invention relates to a method for predicting faults in low-Earth orbit satellite terminals based on cloud-edge collaboration, comprising: S1 transforms the multi-source heterogeneous raw data generated by the low-orbit satellite terminal into a standardized time-series feature matrix that can be directly processed by the neural network model; S2 constructs a dynamic spatiotemporal graph neural network model as a teacher model for capturing the complex spatiotemporal correlation features among terminals across the entire network; S3 transforms the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes the complex graph attention layer and retains only a lightweight temporal convolutional network. The lightweight student model is trained by minimizing the difference between the output distribution of the student model and the teacher model. S4. The low-Earth orbit satellite terminal loads the quantized lightweight student model and performs real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of type C failure If the low-orbit satellite terminal adopts a feature importance evaluation method based on attention weight to select the Q feature indices that result in the highest importance score to support cloud-based interpretable diagnosis. The value of Q is an integer in the range of [1, M], and M is the total dimension of the terminal input features. S5. Cloud-edge collaborative interaction: Define the data interaction protocol between the low-orbit satellite terminal and the ground operation and control center. The event-driven mechanism is adopted so that the low-orbit satellite terminal starts uplink transmission only when the warning is triggered by S4. The low-orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. S6. At the ground operations control center, domain knowledge graphs are used to transform data-driven prediction results into interpretable diagnostic reports.

[0007] To achieve the above objectives, in a second aspect, the present invention relates to a low-Earth orbit satellite terminal fault prediction system based on cloud-edge collaboration, comprising: The feature construction module is used to transform the multi-source heterogeneous raw data generated by the low-orbit satellite terminal into a standardized time-series feature matrix that can be directly processed by the neural network model; A module for constructing a dynamic spatiotemporal graph neural network model is used to build a dynamic spatiotemporal graph neural network model as a high-precision teacher model for capturing complex spatiotemporal correlation features among terminals across the entire network. The model distillation and compression module is used to transform the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes complex graph attention layers and retains only a lightweight temporal convolutional network. The student model is trained by minimizing the difference between the output distribution of the student model and the teacher model. The terminal-side lightweight inference module is used to load the quantized lightweight student model onto the low-Earth orbit satellite terminal for real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of a certain type of failure If the low-orbit satellite is selected, a local early warning will be triggered. The low-orbit satellite will use a feature importance assessment method based on attention weight to select the Q feature indices with the highest importance scores to support interpretable diagnosis in the cloud. The cloud-edge collaborative interaction module is used to define the data interaction protocol between the low-orbit satellite terminal and the ground operation and control center. It adopts an event-driven mechanism so that the low-orbit satellite terminal starts uplink transmission only when the terminal-side lightweight inference module triggers an early warning. The low-orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. An interpretable diagnostic generation module is used in ground operations control centers to transform data-driven prediction results into interpretable diagnostic reports using domain knowledge graphs.

[0008] To achieve the above objectives, in a third aspect, the present invention also relates to a computer-readable storage medium storing instructions that, when executed, perform the above-described method for predicting faults in low-Earth orbit satellite terminals based on cloud-edge collaboration.

[0009] The present invention relates to a method, system, and medium for predicting faults in low-Earth orbit satellite terminals based on cloud-edge collaboration, which has the following advantages compared to existing technologies: The technical solution proposed in this application can effectively overcome the shortcomings of existing low-orbit satellite terminal fault prediction technologies, significantly improve the accuracy, real-time performance and interpretability of fault prediction, and provide strong support for operation and maintenance decisions.

[0010] This application constructs a four-category core feature sub-vector system that includes device status, link quality, spatiotemporal motion, and environmental interference. Through spatiotemporal alignment, Z-score normalization, and sliding window sequence construction, multi-source heterogeneous raw data is transformed into a standardized temporal feature matrix that can be directly processed by the model.

[0011] This application constructs a cloud-based dynamic spatiotemporal graph neural network model, explicitly models the dynamic correlation between terminals caused by satellite switching and space environment coupling, captures the complex spatiotemporal correlation characteristics between terminals across the entire network, and realizes collaborative prediction based on network topology.

[0012] The knowledge graph-driven reasoning module of this application deeply integrates model prediction results with knowledge in fields such as equipment principles and fault propagation paths, automatically generates fault diagnosis paths with causal relationships, and provides interpretable decision support.

[0013] This application designs a knowledge transfer path from a large cloud model to a lightweight terminal model, and comprehensively utilizes knowledge distillation and integer quantization techniques to reduce terminal computing overhead and storage usage while maintaining prediction accuracy. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the low-orbit satellite terminal fault prediction method based on cloud-edge collaboration in Example 1.

[0015] Figure 2A schematic diagram of the structure of the low-orbit satellite terminal fault prediction system based on cloud-edge collaboration in Embodiment 2 of the present invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention and not the entire structure.

[0017] Example 1 For a cloud-edge collaborative method for predicting faults in low-Earth orbit satellite terminals, please refer to [link / reference]. Figure 1-2 As shown, the present invention provides a low-Earth orbit satellite terminal fault prediction method based on cloud-edge collaboration, which is executed by a large-scale constellation network including multiple low-Earth orbit satellite terminals, a ground operation and control center and the cloud, and includes the following steps: S101 to S106.

[0018] S101 transforms the multi-source heterogeneous raw data generated by low-orbit satellite terminals into a standardized time-series feature matrix that can be directly processed by a neural network model.

[0019] Specifically, S101 includes: For low-Earth orbit satellite terminals, the first... The aforementioned low-orbit satellite terminal at time The The feature vector of the class data is The collected raw data is summarized into the following four core feature sub-vectors: device status vector Link quality vector Spatiotemporal motion vector and environmental disturbance vectors The four types of feature sub-vectors mentioned above are concatenated to form the time step. The original multidimensional feature vector As the multi-source heterogeneous raw data; wherein, the device state vector It covers all time-series status data generated within the terminal, including terminal hardware health characteristics (voltage, current, temperature, etc.), software running status (CPU / memory utilization) and internal event counts (error log counts); (2) Link quality vector The timing performance data covering the entire protocol stack of the communication link includes physical layer indicators, protocol layer indicators and service layer indicators; (3) Spatiotemporal motion vectors This includes the relative motion characteristics between the satellite and the ground (satellite elevation angle, azimuth angle, satellite-to-ground distance, Doppler shift) and the terminal's own motion characteristics (if the terminal moves, this includes its GPS position, velocity, and acceleration). (4) Environmental interference vector This includes time-series observation data covering uncontrollable external factors, including characteristics of external environmental influences (ionospheric scintillation, rainfall attenuation, adjacent channel interference power) and electromagnetic environment spectrum data. The four types of feature sub-vectors are concatenated to form the time series data. The original multidimensional feature vector : (1) in, This represents a vector concatenation operation, where M is the total dimension of the features.

[0020] Because the sampling frequencies of the multi-source data are different, a unified time base is needed, requiring spatiotemporal alignment and cleaning: setting a unified time granularity. The original multidimensional feature vector is transformed using linear interpolation. Various subvectors Mapping to a unified timestamp To eliminate dimensional differences and adapt to subsequent neural network inputs, standardization and serialization operations are performed to convert the constructed input sequence at time t. The data is uploaded to the ground control center via satellite link and used as input node features for the subsequent cloud-based spatiotemporal graph neural network, which is then transformed into a standardized time-series feature matrix that can be directly processed by the neural network model.

[0021] Standardization and serialization operations are performed, specifically: first, Z-score standardization is applied to each feature dimension to eliminate the influence of dimensions. Let the mean and standard deviation of the k-th feature in the training set be respectively... and The standardized eigenvalues ,in This is a local constant to prevent the denominator from being zero. The standardized multidimensional eigenvector is represented as follows: Secondly, a time-series sliding window sequence is constructed, with a window length of L and a step size of S. For the i-th terminal, the input sequence at time t is defined as... The terminal will be built The data is uploaded to the ground control center via satellite link and used as input node features for the subsequent cloud-based spatiotemporal graph neural network.

[0022] S102 constructs a dynamic spatiotemporal graph neural network model as a teacher model for capturing the complex spatiotemporal correlation features among terminals across the entire network.

[0023] Specifically, S102 includes: S121 abstracts the N low-Earth orbit satellite terminals within the monitoring range as graph nodes, captures the dynamic evolution characteristics of the low-Earth orbit satellite network topology, and constructs a dynamic adjacency matrix sequence, whose three-dimensional tensor... The three-dimensional tensor is formed by stacking two-dimensional adjacency matrices along the time dimension over L time points within the window period. For the prediction time t, the corresponding time window is... At any time step within this time window The corresponding two-dimensional adjacency matrix can be obtained by slicing from this tensor. ; elements The calculation formula is as follows: (2) in, Iterate through all time steps within the sliding window. Indicates time The unique identifier of the satellite serving the i-th low-Earth orbit satellite terminal. For a moment terminal With terminal The geographic distance between the sub-satellite points, This is the distance attenuation coefficient. and It is a weight hyperparameter.

[0024] The cloud collects the feature sequences of all low-Earth orbit satellite terminals at time t. Furthermore, these features are stacked along the node dimension to construct a tensor of the entire network's node feature sequences. ; S122 Spatiotemporal Feature Fusion Encoding: The teacher model uses a spatiotemporal graph attention network (ST-GNN) for feature extraction. The input to the model is the node feature sequence aggregated in the cloud. and adjacency matrix sequence ; Spatial attention convergence: for each time step within the window Extracting tensors The Time slice This serves as the node feature input for this step, and utilizes the corresponding two-dimensional adjacency matrix at that time. Spatial feature aggregation is performed to obtain the spatially aggregated feature sequence. Node i at time... Space Hidden State The calculation is as follows: (3) in, For nodes The neighborhood group, It is the first The learnable linear transformation matrix of a layer network This is for node j at time [time]. The l-th layer input features, It is a non-linear activation function. The attention coefficient represents the importance of node j to node i, and is calculated and normalized through a shared attention mechanism: (4) in, , It is a learnable attention vector, and LeakyReLU is a modified linear unit function with leakage. For the first low-orbit satellite terminal at time -1 is a spatiotemporal comprehensive representation.

[0025] Time series encoding: spatially aggregated feature sequences The input is fed into the Long Short-Term Memory (LSTM) network of the time coding module to calculate the final spatiotemporal hidden state at time t, which serves as the spatiotemporal comprehensive representation of the i-th low-orbit satellite terminal at time t. Let the cell state of LSTM be... Hidden state is Then the final spacetime hidden state at time t The calculation is as follows: (5) in, This represents standard long short-term memory network unit operations. This indicates that the LSTM is at time step The hidden state, This indicates that the LSTM is at time step The state of cells, Let node i be at time i -1 is a hidden space state; The teacher model represents the first Each terminal at the predicted time The final spatiotemporal comprehensive representation.

[0026] S123 sets the training objective of the teacher model: the training loss function of the teacher model. Losses classified by fault and temporal reconstruction loss composition: (6) in, As a balancing factor, the classification loss uses cross-entropy loss, and the reconstruction loss uses mean squared error.

[0027] S103 transforms the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes the complex graph attention layer and retains only a lightweight temporal convolutional network. The lightweight student model is trained by minimizing the difference between the output distribution of the student model and the teacher model.

[0028] Specifically, S103 includes S131-S133: S131 assumes that the original score vector output by the last fully connected layer of the student model is... The original score vector output by the teacher model is C represents the total number of fault categories.

[0029] S132 trains the student model by minimizing the difference between the output distributions of the student model and the teacher model, including: Define with temperature coefficient The probability distribution calculation function, for any original fraction vector The probability formula for class c is: (7) Where T is the temperature coefficient, and based on the above definition, the softening probability distribution of the c-th type of teacher model is... The softening probability distribution of the c-th student model They are respectively determined as: (8) in ; Define distillation loss function for: (9) in, It is the student model to the real labels Cross-entropy loss, Output the Kullback-Leibler divergence between the probability distributions of the teacher model and the student model. and It is the balance coefficient. It is a scaling factor that balances the magnitude of the gradient; S133 performs knowledge distillation training. After training, the weight parameters of the student model are post-training quantized, converting the 32-bit floating-point weights... Mapped to 8-bit integer weights The quantized model is then sent to a low-orbit satellite terminal via a satellite link.

[0030] (10) in, It is the quantization step size. It is a rounding function. This is the truncation function. The volume compression ratio of the quantized model is approximately: The quantized model is then sent to the terminal via a satellite link.

[0031] S104 The low-Earth orbit satellite terminal loads the quantized lightweight student model and performs real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of type C failure If the low-orbit satellite terminal adopts a feature importance assessment method based on attention weight to select the Q feature indices that result in the highest importance score to support cloud-based interpretable diagnosis, where Q is an integer in the range of [1, M] and M is the total dimension of the terminal input features.

[0032] Specifically, an attention-weight-based feature importance evaluation method is adopted, selecting the Q feature indices with the highest importance scores. The attention weight of input feature k in the student model is given by: Its feature importance score for: (11) Select the Q feature indices with the highest importance scores, denoted as .in, It is an index in the time dimension, where M is the total dimension of the multi-source fusion features of a single low-Earth orbit satellite terminal at each time step. .

[0033] S105 Cloud-Edge Collaborative Interaction: Defines the data interaction protocol between the low-Earth orbit satellite terminal and the ground operation and control center. An event-driven mechanism is adopted so that the low-Earth orbit satellite terminal starts uplink transmission only when S104 triggers an early warning. The low-Earth orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud.

[0034] In S105, the terminal initiates uplink transmission only when an early warning is triggered in S104. Specifically, the data packet structure is as follows: (12) in, It is the unique identifier of the terminal, and TS is the warning timestamp. It is a failure probability vector. It is the key feature index that leads to the warning. The original data fragments within a time window before and after triggering the warning are used for cloud verification. The terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. If no response is received within the specified time, the terminal executes the default security policy.

[0035] S106 utilizes domain knowledge graphs at the ground operations control center to transform data-driven prediction results into interpretable diagnostic reports.

[0036] In S106, a knowledge graph for satellite terminal fault domain is constructed. ,in, It is a collection of entities that includes hardware components, fault phenomena, environmental factors, and root causes. It is a set of relations that includes at least "cause", "accompany", "located in", and "affected by", in order to index the key features uploaded by the terminal. Mapping to fault phenomenon entities in the knowledge graph, define the mapping function: For example, the feature "voltage fluctuation" can be mapped to the entity "abnormal power supply voltage". Starting from the mapped phenomenon entity, a path pointing to the root cause entity is searched in the knowledge graph.

[0037] The shortest path algorithm using confidence-weighted methods is as follows: Let the path be... Representation of knowledge graph A candidate diagnostic path from the entity exhibiting the fault to the entity exhibiting the root cause. ,in The fault phenomenon entity is obtained by mapping the key feature index. As an intermediate entity, Candidate root cause entities; ,in Represents the first [unit] between adjacent entities Relationship edge; This represents the number of relation edges contained in the path.

[0038] The confidence score for the path is calculated as follows: (13) in, It is a knowledge graph relation set The relation edges in the middle, i.e. It is a relationship The prior confidence level is derived from expert experience or historical data statistics. The B paths with the highest confidence scores are selected as candidate diagnostic paths, where B is the preset number of candidate diagnostic paths, satisfying the condition... , Given a set of candidate diagnostic paths, and combining the prediction results with the inference paths (where N is an integer), a structured diagnostic report is generated. :

[0039] in, This represents the prediction result, where, The model represents the first Predictive output for fault types; Indicates diagnostic evidence, It is a set of key feature indexes uploaded by the terminal side, that is, key features selected based on feature importance scores; Indicates the reasoning path, where It is the optimal path selected from the candidate paths; This indicates repair recommendations targeting the root cause, among which, Represents the root cause obtained through reasoning in a knowledge graph. For the root cause retrieved from the map The report provides maintenance recommendations and can be displayed on the terminal or pushed directly to maintenance personnel.

[0040] Example 2 A cloud-edge collaborative low-Earth orbit (LEO) satellite terminal fault prediction system is provided for LEO satellite terminal fault prediction. It is implemented using a large-scale constellation network comprising multiple LEO satellite terminals, a ground control center, and a cloud platform. For implementation details in this example, please refer to [link to relevant documentation]. Figure 2 It includes a feature construction module 61, a dynamic spatiotemporal graph neural network model construction module 62, a model distillation and compression module 63, a terminal-side lightweight inference module 64, a cloud-edge collaborative interaction module 65, and an interpretable diagnostic generation module 66.

[0041] The feature construction module 61 is used to transform the multi-source heterogeneous raw data generated by the low-orbit satellite terminal into a standardized time-series feature matrix that can be directly processed by the neural network model.

[0042] The module 62 for constructing a dynamic spatiotemporal graph neural network model is used to construct a dynamic spatiotemporal graph neural network model as a high-precision teacher model for capturing complex spatiotemporal correlation features among terminals across the entire network.

[0043] The model distillation and compression module 63 is used to convert the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes complex graph attention layers and retains only a lightweight temporal convolutional network. The student model is trained by minimizing the difference between the output distribution of the student model and the teacher model.

[0044] Terminal-side lightweight inference module 64: used by the low-Earth orbit satellite terminal to load the quantized lightweight student model and perform real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of a certain type of failure If the low-orbit satellite finally adopts the feature importance assessment method based on attention weight to select the Q feature indices with the highest importance scores to support cloud-based interpretable diagnosis. The value of Q is an integer in the range of [1, M], and M is the total dimension of the terminal input features.

[0045] The cloud-edge collaborative interaction module 65 is used to define the data interaction protocol between the low-orbit satellite terminal and the ground operation and control center. It adopts an event-driven mechanism so that the low-orbit satellite terminal starts uplink transmission only when the lightweight inference module on the terminal side triggers an early warning. The low-orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud.

[0046] The interpretable diagnostic generation module 66 is used in the ground operations control center to transform data-driven prediction results into interpretable diagnostic reports using domain knowledge graphs.

[0047] The cloud-edge collaborative low-Earth orbit satellite terminal fault prediction system of this embodiment is the same as the cloud-edge collaborative low-Earth orbit satellite terminal fault prediction method described in Embodiment 1 in terms of implementation process, method and effect, and will not be repeated here.

[0048] Example 3 This invention relates to a computer-readable storage medium storing instructions. When the instructions are executed, they perform a cloud-edge collaborative low-Earth orbit satellite terminal fault prediction method according to Embodiment 1. The execution process and effects are the same as those described in Embodiment 1, and will not be repeated here.

[0049] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0050] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting faults in low-Earth orbit satellite terminals based on cloud-edge collaboration, characterized in that, include: S1 transforms the multi-source heterogeneous raw data generated by the low-orbit satellite terminal into a standardized time-series feature matrix that can be directly processed by the neural network model; S2 constructs a dynamic spatiotemporal graph neural network model as a teacher model for capturing the complex spatiotemporal correlation features among terminals across the entire network; S3 transforms the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes the complex graph attention layer and retains only a lightweight temporal convolutional network. The lightweight student model is trained by minimizing the difference between the output distribution of the student model and the teacher model. S4 The low-Earth orbit satellite terminal loads the quantized lightweight student model and performs real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of a certain type of failure If the low-orbit satellite terminal adopts a feature importance evaluation method based on attention weight to select the Q feature indices that result in the highest importance score to support cloud-based interpretable diagnosis. The value of Q is an integer in the range of [1, M], and M is the total dimension of the terminal input features. S5 Cloud-Edge Collaborative Interaction: Defines the data interaction protocol between the low-Earth orbit satellite terminal and the ground operation and control center. An event-driven mechanism is adopted so that the low-Earth orbit satellite terminal starts uplink transmission only when S4 triggers an early warning. The low-Earth orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. S6 utilizes domain knowledge graphs in the ground operations control center to transform data-driven prediction results into interpretable diagnostic reports.

2. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 1, characterized in that, S1 specifically includes: For low-Earth orbit satellite terminals, the first... The aforementioned low-orbit satellite terminal at time The The feature vector of the class data is The collected raw data is summarized into the following four core feature sub-vectors: device status vector Link quality vector Spatiotemporal motion vector and environmental disturbance vectors The four types of feature sub-vectors mentioned above are concatenated to form the time step. The original multidimensional feature vector As the aforementioned multi-source heterogeneous raw data; in, This represents a vector concatenation operation, where M is the total dimension of the features; Spatiotemporal alignment and cleaning: setting a uniform time granularity The original multidimensional feature vector is transformed using linear interpolation. Various subvectors Mapping to a unified timestamp Above, perform standardization and serialization operations to construct the input sequence at time t. The data is uploaded to the ground control center via satellite link and used as input node features for the subsequent cloud-based spatiotemporal graph neural network, which is then transformed into a standardized time-series feature matrix that can be directly processed by the neural network model.

3. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 2, characterized in that, S2 specifically includes: The N low-Earth orbit satellite terminals within the monitoring range are abstracted as graph nodes. The dynamic evolution characteristics of the low-Earth orbit satellite network topology are captured, and a dynamic adjacency matrix sequence is constructed, whose three-dimensional tensor... The three-dimensional tensor is formed by stacking two-dimensional adjacency matrices along the time dimension over L time points within the window period. For the prediction time t, the corresponding time window is... At any time step within this time window The corresponding two-dimensional adjacency matrix can be obtained by slicing from this tensor. ; Cloud-based collection of monitoring data The standardized feature sequence of a low-Earth orbit satellite terminal within the sliding window corresponding to time t And stack them along the node dimension to construct a tensor of the feature sequence of all nodes in the network. ,in, Indicates the number of terminal nodes. Indicates the length of the sliding window. This represents the total dimension of features at each time step. This tensor serves as the node feature input for the cloud-based dynamic spatiotemporal graph neural network, used to characterize the state evolution of all network terminals in the time dimension. Spatiotemporal feature fusion encoding: The teacher model uses a spatiotemporal graph attention network (ST-GNN) for feature extraction. The input to the model is the node feature sequence aggregated in the cloud. and adjacency matrix sequence ; Spatial attention convergence: for each time step within the window Extracting tensors The Time slice The node features are used as input for this step, and the corresponding two-dimensional adjacency matrix at this time is utilized. Spatial feature aggregation is performed to obtain the spatially aggregated feature sequence; Time series encoding: spatially aggregated feature sequences The input is fed into the Long Short-Term Memory (LSTM) network of the time coding module to calculate the final spatiotemporal hidden state at time t, which serves as the spatiotemporal comprehensive representation of the i-th low-orbit satellite terminal at time t. Set the training objective for the teacher model: the training loss function of the teacher model. Losses classified by fault and temporal reconstruction loss composition: ； in, As a balancing factor, the classification loss uses cross-entropy loss, and the reconstruction loss uses mean squared error.

4. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 3, characterized in that: The slices in this tensor yield the corresponding two-dimensional adjacency matrix. Specifically: elements The calculation formula is as follows: ； in, Iterate through all time steps within the sliding window. Indicates time The unique identifier of the satellite serving the i-th low-Earth orbit satellite terminal. For a moment The geographic distance between the i-th and j-th low-Earth orbit satellite terminals. , This indicates the number of low-Earth orbit satellite terminals within the monitoring range. This is the distance attenuation coefficient. and It is a weight hyperparameter; The two-dimensional adjacency matrix corresponding to this moment is used. Spatial feature aggregation is performed to obtain the spatially aggregated feature sequence, specifically: node i at time Space Hidden State The calculation is as follows: ； in, For nodes The neighborhood group, It is the first The learnable linear transformation matrix of a layer network This is for node j at time [time]. The l-th layer input features, It is a non-linear activation function. Let be the attention coefficient, representing the importance of node j to node i, which is calculated and normalized through the shared attention mechanism; The final spatiotemporal hidden state at time t is calculated as follows: Let the cell state of LSTM be... Hidden state is Then the final spacetime hidden state at time t ; The calculation is as follows: in, This represents standard long short-term memory network unit operations. This indicates that the LSTM is at time step The hidden state, This indicates that the LSTM is at time step The state of cells, Let node i be at time i -1 is a hidden space state; The teacher model represents the first Each terminal at the predicted time The final spatiotemporal comprehensive representation.

5. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 3, characterized in that, S3 specifically refers to: Let the original score vector output by the last fully connected layer of the student model be... The original score vector output by the teacher model is C represents the total number of fault categories; The method of training the student model by minimizing the difference between the output distributions of the student model and the teacher model includes: Defined with temperature coefficient The probability distribution calculation function, for any original fraction vector The probability calculation formula for class c is as follows: Where T is the temperature coefficient, and based on the above definition, the softening probability distribution of the c-th type of teacher model is... The softening probability distribution of the c-th student model They are respectively determined as: ,in ; Define distillation loss function for: ） in, It is the student model to the real labels Cross-entropy loss, Output the Kullback-Leibler divergence between the probability distributions of the teacher model and the student model. and It is the balance coefficient. It is a scaling factor that balances the magnitude of the gradient; Knowledge distillation training is performed. After training, the weight parameters of the student model are post-training quantized, converting the 32-bit floating-point weights... Mapped to 8-bit integer weights The quantized model is then sent to a low-orbit satellite terminal via a satellite link.

6. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 3 is characterized in that, In S4, a feature importance evaluation method based on attention weights is adopted, selecting the Q feature indices with the highest importance scores. Specifically, let the attention weight of input feature k in the student model be... Its feature importance score for: ； Select the Q feature indices with the highest importance scores, denoted as ; in, It is an index in the time dimension, where M is the total dimension of the multi-source fusion features of a single low-Earth orbit satellite terminal at each time step. .

7. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 3, characterized in that, In S5, the terminal initiates uplink transmission only when an alert is triggered by S4. Specifically, the data packet structure is as follows: ； in, It is the unique identifier of the terminal, and TS is the warning timestamp. It is a failure probability vector. It is the key feature index that leads to the warning. The original data fragments within a time window before and after triggering the warning are used for cloud verification. The terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. If no response is received within the specified time, the terminal executes the default security policy.

8. The low-orbit satellite terminal fault prediction method based on cloud-edge collaboration according to claim 3 is characterized in that, In S6, a knowledge graph for satellite terminal fault domains is constructed. ,in, It is a collection of entities that includes hardware components, fault phenomena, environmental factors, and root causes. It is a set of relations that includes at least "cause", "accompany", "located in", and "affected by", in order to index the key features uploaded by the terminal. Mapping to fault phenomenon entities in the knowledge graph, define the mapping function: ; The shortest path algorithm using confidence-weighted methods is as follows: Let the path be... Representation of knowledge graph A candidate diagnostic path from the entity exhibiting the fault to the entity exhibiting the root cause. ,in The fault phenomenon entity is obtained by mapping the key feature index. As an intermediate entity, Candidate root cause entities; ,in Represents the first [unit] between adjacent entities Relationship edge; This represents the number of relation edges contained in the path. The confidence score for the path is calculated as follows: ； in, It is a knowledge graph relation set The relation edges in the middle, i.e. It is a relationship The prior confidence level is derived from expert experience or historical data statistics. The B paths with the highest confidence scores are selected as candidate diagnostic paths, where B is the preset number of candidate diagnostic paths, satisfying the condition... , A set of candidate diagnostic paths is generated; by combining the prediction results and the inference path, a structured diagnostic report is produced. : ； in, This represents the prediction result, where, The model represents the first Predictive output for fault types; Indicates diagnostic evidence, It is a set of key feature indexes uploaded by the terminal side, that is, key features selected based on feature importance scores; Indicates the reasoning path, where It is the optimal path selected from the candidate paths; This indicates repair recommendations targeting the root cause, among which, This represents the root cause entity obtained through reasoning in the knowledge graph. For the root cause retrieved from the map The report provides maintenance recommendations and can be displayed on the terminal or pushed directly to maintenance personnel.

9. A low-orbit satellite terminal fault prediction system based on cloud-edge collaboration, characterized in that, include: The feature construction module is used to transform the multi-source heterogeneous raw data generated by the low-orbit satellite terminal into a standardized time-series feature matrix that can be directly processed by the neural network model; A module for constructing a dynamic spatiotemporal graph neural network model is used to build a dynamic spatiotemporal graph neural network model as a high-precision teacher model for capturing complex spatiotemporal correlation features among terminals across the entire network. The model distillation and compression module is used to transform the teacher model into a lightweight student model suitable for deployment on the low-Earth orbit satellite terminal. The student model removes complex graph attention layers and retains only a lightweight temporal convolutional network. The student model is trained by minimizing the difference between the output distribution of the student model and the teacher model. The terminal-side lightweight inference module is used to load the quantized lightweight student model onto the low-Earth orbit satellite terminal for real-time online inference; the low-Earth orbit satellite terminal collects data in real time and constructs the input sequence. After inputting the lightweight student model, a fault probability vector is obtained. Where C represents the number of fault categories, and a warning threshold is set. If the probability of Class C failure If the low-orbit satellite is selected, a local early warning will be triggered. The low-orbit satellite will use a feature importance assessment method based on attention weight to select the Q feature indices with the highest importance scores to support interpretable diagnosis in the cloud. The cloud-edge collaborative interaction module is used to define the data interaction protocol between the low-orbit satellite terminal and the ground operation and control center. It adopts an event-driven mechanism so that the low-orbit satellite terminal starts uplink transmission only when the terminal-side lightweight inference module triggers an early warning. The low-orbit satellite terminal listens to the downlink channel and receives the diagnostic report ID returned by the cloud. An interpretable diagnostic generation module is used in ground operations control centers to transform data-driven prediction results into interpretable diagnostic reports using domain knowledge graphs.

10. A computer-readable storage medium, characterized in that: The storage medium stores instructions that, when executed, perform the low-orbit satellite terminal fault prediction method based on cloud-edge collaboration as described in any one of claims 1-8.

Images