Satellite communication fault root cause analysis method based on size model cooperation
By employing a fault root cause analysis method that combines large and small models, and integrating large models with uncertainty estimation small models, the challenge of fault root cause analysis in large-scale satellite internet systems has been solved. This method enables efficient identification and localization of known and unknown faults, and enhances analytical capabilities in complex fault scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 54TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION
- Filing Date
- 2026-05-11
- Publication Date
- 2026-07-31
AI Technical Summary
Existing satellite fault detection methods are insufficient to cover all fault ranges in large-scale satellite internet systems. Traditional small models rely on human experience, which is inefficient, and it is difficult to locate unknown faults in complex topological dependencies.
A fault root cause analysis method that combines large and small models is adopted. It combines a large model-driven root cause analysis model with an uncertainty estimation fault prediction small model. Through knowledge graph and graph attention feature extraction, it achieves high-precision prediction of fault state and identification of unknown anomalies.
It can simultaneously cover the identification of known faults and the root cause analysis of unknown faults, improve the root cause analysis effect in complex fault scenarios, enhance the identification capability of key abnormal nodes, and is suitable for complex dynamic scenarios of real satellite communication networks.
Smart Images
Figure CN122496084A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a satellite communication fault root cause analysis method based on a combination of large and small models in the field of satellite communication, which is particularly suitable for fault root cause analysis of large-scale satellite internet systems. Background Technology
[0002] With the rapid development of next-generation satellite communication technologies, especially the rapid construction of integrated high- and low-Earth orbit satellite internet systems, the constellation size, number of nodes, and service types of satellite communication networks are continuously increasing. Furthermore, the coupling between network nodes, payload equipment, link relationships, and service tasks is constantly improving, leading to fault patterns characterized by multiple sources, multiple layers, dynamic changes, and strong correlations. Failure to detect anomalies promptly and accurately pinpoint the root cause of faults can result in network performance degradation, communication service interruptions, and inability to respond quickly in special mission scenarios.
[0003] Existing satellite fault detection methods typically employ analytical models, SVMs, random forests, deep neural networks, and time-series modeling to classify and identify known fault samples. While these methods perform well in identifying faults within the training distribution, for large-scale satellite internet constellations, the types of faults are numerous and their distribution complex, making it difficult to cover the entire range of faults using traditional methods alone. When anomalies not covered during the training phase occur, traditional models will still assign a known category, failing to effectively indicate that the sample exceeds existing cognitive boundaries.
[0004] On the other hand, existing satellite communication operation and maintenance systems generally rely on expert experience, fault manuals, and preset rules for root cause analysis. This approach is highly dependent on accumulated human knowledge and suffers from low efficiency, insufficient generalization ability, and difficulty in locating unknown faults when dealing with multi-source heterogeneous state data, complex topological dependencies, and cross-level fault chains in high-Earth orbit integrated networks. Especially in the large-scale nodes of satellite internet, locating specific node faults is quite difficult, and traditional small models or rule-based methods are insufficient to meet practical needs. However, the rapidly developing large model technology has shown strong advantages in node state analysis and tracing. Summary of the Invention
[0005] This invention proposes a satellite communication fault root cause analysis method that combines large and small models to address the challenges of exhaustively listing fault states, obtaining complete training samples, and locating root causes in complex and abnormal scenarios within satellite communication networks. In the field of satellite communication operation and maintenance management, this invention collaboratively designs a large-model-driven root cause analysis model with an uncertainty estimation fault prediction small model. Furthermore, it integrates knowledge graphs, graph attention feature extraction, and large-model inference to achieve fault root cause localization. This enables the system to accurately predict known fault states and identify unknown abnormal states outside the training distribution.
[0006] The present invention adopts the following technical solution:
[0007] A method for root cause analysis of satellite communication failures using a combination of large and small models includes the following steps:
[0008] Step 1: Perform unified information representation on multi-source heterogeneous data in the satellite communication system to obtain information representation data for fault prediction and root cause analysis, and construct a knowledge graph of the satellite communication system nodes;
[0009] Step 2: Construct a small model for uncertainty estimation and fault prediction. Use the fault prediction information representation data as input to obtain the fault prediction result. Determine whether a fault exists based on the fault prediction result. If so, continue to Step 3; otherwise, end the entire process.
[0010] Step 3: Construct a large model, including a graph attention feature extraction module for root cause analysis, a feature representation layer, a Transformer module, a feature encoder, and a decoding layer;
[0011] Step 4: The graph attention feature extraction module extracts features based on the graph topology of the knowledge graph and the information representation data from root cause analysis to obtain node information features; the feature representation layer calculates the feature embedding of state information based on the information representation data from root cause analysis.
[0012] Step 5: Embed the features of the state information with the features of the node information, extract the graph features through the Transformer module, and calculate the graph-level vector with the state of the satellite communication network topology.
[0013] Step 6: Further feature representation is performed on the graph-level vectors through a feature encoder constructed with multiple fully connected layers to complete the fusion of graph features and state features, and obtain the final representation vector;
[0014] Step 7: Input the final representation vector into the decoding layer to locate the root cause of satellite communication faults and obtain the root cause analysis results.
[0015] Furthermore, in step 1, the multi-source heterogeneous data includes satellite payload telemetry data, link status data, inter-satellite and satellite-to-ground connection relationships, service task status, equipment attribute information, fault log information, and operation and maintenance rule text.
[0016] The constructed knowledge graph of satellite communication system nodes is as follows:
[0017]
[0018] In the formula, Represents a set of nodes. Represents a set of edges with relational types. Represents a set of relations;
[0019] in:
[0020] Based on the physical and operational organization of the satellite communication system, the node set is... Divided into:
[0021]
[0022] In the formula, Represents the set of constellation nodes. Represents a set of satellite nodes. Represents the set of beam nodes. Represents a set of business task nodes. This represents the set of load devices and board nodes;
[0023] The set of edges with relational types is determined based on the specific connection relationships within the actual constellation system. ;
[0024] Define relation set for:
[0025]
[0026] In the formula, Indicates a subordinate relationship. Indicates a supporting relationship. Indicates the mapping relationship. Indicates a connection relationship. Indicates a constraint relationship;
[0027] To characterize the connectivity properties under different relation types, a relation adjacency tensor is constructed:
[0028]
[0029] In the formula, Represents a node With nodes Does a relationship of type exist between them? The connection.
[0030] Furthermore, step 2 specifically involves:
[0031] A small model for fault prediction based on uncertainty estimation is constructed, which includes a backbone network and an uncertainty network. A deep neural network architecture is used as the backbone network, and the uncertainty network consists of multiple fully connected layers to estimate the uncertainty of the fault prediction results of the backbone network.
[0032] Uncertainty estimation is used to estimate the uncertainty of the information representation data of fault prediction using a small model, and the variance of the current result is obtained. It is then determined whether data drift has occurred in the information representation data of fault prediction. Based on the result, it is determined whether a fault exists. If the variance is greater than or equal to the threshold, step 3 is continued. If the variance is less than the threshold, the entire process ends.
[0033] The graph attention feature extraction module extracts features based on the graph topology of the knowledge graph and the information representation data from root cause analysis to obtain node information features;
[0034] Furthermore, the graph attention feature extraction module in step 4 includes a multi-head attention layer, two cascaded graph convolutional neural networks, a feature cascade layer, and a fully connected layer;
[0035] The multi-head attention layer performs feature aggregation based on the graph topology of the knowledge graph and the information representation data of the root cause analysis. Then, it calculates new node information features through two cascaded graph convolutional neural networks. The new node information features are then concatenated with the node features after feature aggregation through the feature concatenation layer. The fully connected layer recompresses the dimension of the concatenated vector back to the original dimension.
[0036] The splicing is defined as follows:
[0037]
[0038] in, This represents the concatenated vector. This represents the node information features calculated using a graph convolutional neural network. This represents the node features after feature aggregation.
[0039] Furthermore, in step 4, the feature representation layer calculates the feature embedding of the state information based on the information representation data from the root cause analysis. The specific process is as follows:
[0040] The information representation data from the root cause analysis is input into the word segmenter for calculation, resulting in the corresponding position embedding, position index, and mask matrix. The position index corresponds to the index number of each character in the word segmenter, the position embedding is used to identify the position of the character in the sequence, and the mask matrix is used to distinguish between the input and generated content in the position index.
[0041] The feature representation layer, composed of a graph neural network, is used to map the indices of the input positions to obtain... Dense vectors of dimension , This refers to the feature embedding obtained from the state information, where n represents the batch size. The length of the dense vector is used to calculate the detailed identifier of the state information feature embedding, which is as follows:
[0042]
[0043]
[0044] in, Indicates a word segmenter. This represents the information characterization data used in root cause analysis. These are the position index, position embedding, and mask matrix of the state information, respectively. This represents the state information feature embedding obtained after the feature representation layer. Forward reasoning represents the feature representation layer.
[0045] Furthermore, in step 5, the graph-level vector is calculated through the following process:
[0046]
[0047]
[0048]
[0049] in, Represents node feature embedding. This represents the graph-level vector after attention calculation, Linear(.) represents a fully connected layer, and T represents the transpose of the matrix. This is the scaling factor in the attention mechanism.
[0050] Furthermore, step 6 specifically involves:
[0051] The calculated and The features are summed and then passed through multiple fully connected layers to construct a feature encoder, resulting in a feature embedding that simultaneously contains state information and information from graph-level vectors. This is the final representation vector.
[0052] Furthermore, step 7 specifically involves:
[0053] Feature embedding The input decoding layer is used to locate the root cause of satellite communication faults, and the root cause analysis results are obtained:
[0054]
[0055] in, The results of the root cause analysis should include at least one of the following: the faulty node, the cause of the fault, the associated equipment, the affected links, the fault propagation path, and the recommended objects to be investigated.
[0056] Compared with the prior art, the present invention has the following beneficial effects:
[0057] This invention combines a small uncertainty estimation model with a large-scale root cause analysis model, simultaneously covering both known fault identification and unknown fault root cause analysis scenarios. This overcomes the limitation of traditional small models in addressing the root cause analysis of node faults in large-scale satellite communication systems. Furthermore, this invention introduces graph attention feature extraction and feature concatenation mechanisms into the RGCN structure, enabling it to learn graph topology propagation information and inherit original node attribute features, thereby improving the ability to identify key abnormal nodes during the root cause analysis phase. By embedding graph-level vector encoding into the large-scale model's forward propagation process, the large-scale model can perform comprehensive reasoning by combining network structure information, log information, and state semantics, significantly improving the root cause analysis performance in complex fault scenarios. This invention does not require pre-enumerating all fault states or collecting separate training data for each type of unknown fault, making it more suitable for application in complex dynamic scenarios in real satellite communication networks. Attached Figure Description
[0058] Figure 1 This is a schematic diagram of a satellite communication fault root cause analysis method based on a combination of large and small models, according to an embodiment of the present invention. Detailed Implementation
[0059] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.
[0060] In one embodiment, such as Figure 1 As shown, this invention provides a satellite communication fault root cause analysis method based on a combination of large and small models, which includes the following steps:
[0061] Step 1: Perform unified information representation on multi-source heterogeneous data in the satellite communication system to obtain representation data for fault prediction and root cause analysis, and construct a knowledge graph of satellite communication system nodes;
[0062] The multi-source heterogeneous data includes at least: satellite payload telemetry data, link status data, inter-satellite and satellite-to-ground connection relationships, service mission status, equipment attribute information, fault log information, and operation and maintenance rule text. To enable unified processing of data from different sources, the satellite communication system is abstracted into a heterogeneous knowledge graph:
[0063]
[0064] in, Represents a set of nodes. Represents a set of edges with relational types. This represents a set of relationships. Through the above definition, entity objects, hierarchical relationships, and attribute information in a satellite communication system can be uniformly mapped onto a graph structure, providing standard input for subsequent fault prediction and root cause analysis.
[0065] Furthermore, based on the physical and operational organization of the satellite communication system, the node set is... Divided into:
[0066]
[0067] in, Represents the set of constellation nodes. Represents a set of satellite nodes. Represents the set of beam nodes. Represents a set of business task nodes. This represents the payload equipment and board node set. This layered structure can simultaneously preserve the topological relationships and semantic constraints of the constellation layer, satellite layer, resource layer, and device layer.
[0068] also, The set of edges representing relational relationships needs to be determined based on the specific connections within the actual constellation system. In relational modeling, the relation set is defined as:
[0069]
[0070] in, Indicates a subordinate relationship Indicates a supporting relationship, Represents mapping relationship Indicates connection relationship, Representing constraint relationships. To characterize the connection properties under different relationship types, a relationship adjacency tensor is constructed:
[0071]
[0072] in, Represents a node With nodes Does a relationship of type exist between them? The connections described above. Compared to a single adjacency matrix, the relational adjacency tensor can more accurately express multi-level and multi-semantic dependencies in satellite communication networks.
[0073] Step 2: Construct a small model for uncertainty estimation and fault prediction. Use the fault prediction information representation data as input to obtain the fault prediction result. Determine whether a fault exists based on the fault prediction result. If so, continue to Step 3; otherwise, end the entire process.
[0074] The uncertainty estimation fault prediction mini-model comprises two key modules: a backbone network and an uncertainty network. A typical deep neural network architecture is used as the backbone network, while the uncertainty network is a general uncertainty estimation neural network type, consisting of multiple fully connected layers (FC). This invention leverages its interval estimation characteristic to perform uncertainty estimation on the information representation data for fault prediction, obtaining a variance estimate of the current result, which is used to determine whether data drift has occurred in the information representation data for fault prediction. Subsequently, based on the result, it is determined whether a fault exists. If the variance is greater than or equal to a threshold, step 3 is continued; if the variance is less than the threshold, the entire process ends.
[0075] Step 3: Construct a large model, including a graph attention feature extraction module for root cause analysis, a feature representation layer, a Transformer module, a feature encoder, and a decoding layer;
[0076] Step 4: The graph attention feature extraction module extracts features based on the graph topology of the knowledge graph and the information representation data from root cause analysis to obtain node information features;
[0077] The graph attention feature extraction module includes a multi-head attention layer, two cascaded graph convolutional neural networks, a feature cascade layer, and a fully connected layer;
[0078] The multi-head attention layer performs feature aggregation based on the graph topology of the knowledge graph and the information representation data of the root cause analysis. Then, it calculates new node information features through two cascaded graph convolutional neural networks. The new node information features are then concatenated with the node features after feature aggregation through the feature concatenation layer. The fully connected layer recompresses the dimension of the concatenated vector back to the original dimension.
[0079] The splicing is defined as follows:
[0080]
[0081] in, This represents the concatenated vector. This represents the node information features calculated using a graph convolutional neural network. This represents the node features after feature aggregation.
[0082] The feature representation layer calculates the feature embedding of state information based on the information from root cause analysis; the specific process is as follows:
[0083] The information representation data from the root cause analysis is input into the word segmenter for calculation, resulting in the corresponding position embedding, position index, and mask matrix. The position index corresponds to the index number of each character in the word segmenter, the position embedding is used to identify the position of the character in the sequence, and the mask matrix is used to distinguish between the input and generated content in the position index.
[0084] The feature representation layer, composed of a graph neural network, is used to map the indices of the input positions to obtain... (In this invention) The value is 4096. Too small a value will cause the context to exceed the limit. In this invention patent, n can take values such as 4096, 8192, etc. (depending on the context length limit of the large model). (Dense vector of dimensions) , This refers to the feature embedding obtained from the state information. The detailed identifier of the calculated state information feature embedding is as follows:
[0085]
[0086]
[0087] in, A tokenizer representing large LLM models. This is represented as information characterization data for root cause analysis. These are the position index, position embedding, and mask matrix of the state information, respectively. This represents the state information feature embedding obtained after the feature representation layer. Forward reasoning represents the feature representation layer.
[0088] Step 5: Embed the features of the calculated state information with the node features, extract graph features through the Transformer module, calculate the graph-level vector with the state of the satellite communication network topology, and complete the fusion of graph features and state features;
[0089] The graph-level vectors are calculated through the following process:
[0090]
[0091]
[0092]
[0093] in, Represents node feature embedding. This represents the graph-level vector after attention calculation, Linear(.) represents a fully connected layer, and T represents the transpose of the matrix. This is the scaling factor in the attention mechanism.
[0094] Step 6: Further feature representation is performed on the graph-level vectors through a feature encoder constructed with multiple fully connected layers to complete the fusion of graph features and state features, and obtain the final representation vector;
[0095] The calculated Firstly, in the reasoning process of LLMs, with The summation and processing by a feature encoder yield a feature embedding that simultaneously contains state information and information from the graph-level vectors. Its function is twofold: first, to make it suitable for the input dimensions of LLMs models; and second, to transform it into the semantic space of LLMs, which helps LLMs to achieve root cause analysis through forward propagation of this feature.
[0096] Step 7: Input the final representation vector into the decoding layer to locate the root cause of satellite communication faults and obtain the root cause analysis results; specifically:
[0097] Perform large-model-driven root cause analysis and embed features. The input decoding layer is used to locate the root cause of satellite communication faults, and the root cause analysis results are obtained:
[0098]
[0099] in, The root cause analysis results include at least one of the following: faulty node, fault cause, associated equipment, affected links, fault propagation path, or suggested investigation targets. Since this input already integrates graph structure dependency information from the knowledge graph and task semantic information from the state text, the large model can combine its causal reasoning and contextual understanding capabilities to perform more accurate localization analysis of complex satellite communication faults. The large model in this invention employs large language models with decoders such as the Qwen and Llama series. During the training phase, the original large model parameters are frozen, and only the feature representation layer, attention fusion layer, and feature encoder are incrementally fine-tuned to reduce training costs and maintain the stable semantic capabilities of the original large model. Simultaneously, root cause localization labels are added to the state information template, such as "Please analyze the state logs of each node based on the relationships between nodes in the satellite communication system to determine the node that has failed in the current satellite communication system: {sat_node}", to enhance the model's relevance to the root cause localization task.
[0100] This embodiment proposes a satellite communication fault root cause analysis method that combines a small-scale model with a large-scale model. By integrating a small-scale uncertainty estimation model with a large-scale root cause analysis model, it can simultaneously cover both known fault identification and unknown fault root cause analysis scenarios, overcoming the problem that traditional small models struggle to solve the root cause analysis of node faults in large-scale satellite communication systems. Furthermore, this invention introduces graph attention feature extraction and feature concatenation mechanisms into the RGCN structure, enabling it to learn graph topology propagation information and inherit original node attribute features, thereby improving the ability to identify key abnormal nodes during the root cause analysis stage. By embedding graph-level vector encoding into the large-scale model's forward propagation process, the large-scale model can perform comprehensive reasoning by combining network structure information, log information, and state semantics, significantly improving the root cause analysis performance in complex fault scenarios. This invention does not require pre-enumerating all fault states or collecting separate training data for each type of unknown fault, making it more suitable for application in complex dynamic scenarios in real satellite communication networks.
Claims
1. A method for root cause analysis of satellite communication faults using a combination of large and small models, characterized in that, Includes the following steps: Step 1: Perform unified information representation on multi-source heterogeneous data in the satellite communication system to obtain information representation data for fault prediction and root cause analysis, and construct a knowledge graph of the satellite communication system nodes; Step 2: Construct a small model for uncertainty estimation and fault prediction. Use the fault prediction information representation data as input to obtain the fault prediction result. Determine whether a fault exists based on the fault prediction result. If so, continue to Step 3; otherwise, end the entire process. Step 3: Construct a large model, including a graph attention feature extraction module for root cause analysis, a feature representation layer, a Transformer module, a feature encoder, and a decoding layer; Step 4: The graph attention feature extraction module extracts features based on the graph topology of the knowledge graph and the information representation data from root cause analysis to obtain node information features; The feature representation layer calculates the feature embedding of state information based on the information from root cause analysis. Step 5: Embed the features of the state information with the features of the node information, extract the graph features through the Transformer module, and calculate the graph-level vector with the state of the satellite communication network topology. Step 6: Further feature representation is performed on the graph-level vectors through a feature encoder constructed with multiple fully connected layers to complete the fusion of graph features and state features, and obtain the final representation vector; Step 7: Input the final representation vector into the decoding layer to locate the root cause of satellite communication faults and obtain the root cause analysis results.
2. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 1, characterized in that, In step 1, the multi-source heterogeneous data includes satellite payload telemetry data, link status data, inter-satellite and satellite-to-ground connection relationships, service task status, equipment attribute information, fault log information, and operation and maintenance rule text; The constructed knowledge graph of satellite communication system nodes is as follows: In the formula, Represents a set of nodes. Represents a set of edges with relational types. Represents a set of relations; in: Based on the physical and operational organization of the satellite communication system, the node set is... Divided into: In the formula, Represents the set of constellation nodes. Represents a set of satellite nodes. Represents the set of beam nodes. Represents a set of business task nodes. This represents the set of load devices and board nodes; The set of edges with relational types is determined based on the specific connection relationships within the actual constellation system. ; Define relation set for: In the formula, Indicates a subordinate relationship. Indicates a supporting relationship. Indicates the mapping relationship. Indicates a connection relationship. Indicates a constraint relationship; To characterize the connectivity properties under different relation types, a relation adjacency tensor is constructed: In the formula, Represents a node With nodes Does a relationship of type exist between them? The connection.
3. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 1, characterized in that, Step 2 is as follows: A small model for fault prediction based on uncertainty estimation is constructed, which includes a backbone network and an uncertainty network. A deep neural network architecture is used as the backbone network, and the uncertainty network consists of multiple fully connected layers to estimate the uncertainty of the fault prediction results of the backbone network. Uncertainty estimation is used to estimate the uncertainty of the information representation data of fault prediction using a small model, and the variance of the current result is obtained. It is then determined whether data drift has occurred in the information representation data of fault prediction. Based on the result, it is determined whether a fault exists. If the variance is greater than or equal to the threshold, step 3 is continued. If the variance is less than the threshold, the entire process ends. The graph attention feature extraction module extracts features based on the graph topology of the knowledge graph and the information representation data from root cause analysis, thus obtaining node information features.
4. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 1, characterized in that, The graph attention feature extraction module in step 4 includes a multi-head attention layer, two cascaded graph convolutional neural networks, a feature cascade layer, and a fully connected layer; The multi-head attention layer performs feature aggregation based on the graph topology of the knowledge graph and the information representation data of the root cause analysis. Then, it calculates new node information features through two cascaded graph convolutional neural networks. The new node information features are then concatenated with the node features after feature aggregation through the feature concatenation layer. The fully connected layer recompresses the dimension of the concatenated vector back to the original dimension. The splicing is defined as follows: in, This represents the concatenated vector. This represents the node information features calculated using a graph convolutional neural network. This represents the node features after feature aggregation.
5. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 1, characterized in that, In step 4, the feature representation layer calculates the feature embedding of the state information based on the information from the root cause analysis. The specific process is as follows: The information representation data from the root cause analysis is input into the word segmenter for calculation, resulting in the corresponding position embedding, position index, and mask matrix. The position index corresponds to the index number of each character in the word segmenter, the position embedding is used to identify the position of the character in the sequence, and the mask matrix is used to distinguish between the input and generated content in the position index. The feature representation layer, composed of a graph neural network, is used to map the indices of the input positions to obtain... Dense vectors of dimension , This refers to the feature embedding obtained from the state information, where n represents the batch size. The length of the dense vector is used to calculate the detailed identifier of the state information feature embedding, which is as follows: in, Indicates a word segmenter. This represents the information characterization data used in root cause analysis. These are the position index, position embedding, and mask matrix of the state information, respectively. This represents the state information feature embedding obtained after the feature representation layer. Forward reasoning represents the feature representation layer.
6. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 5, characterized in that, In step 5, the graph-level vector is calculated through the following process: in, Represents node feature embedding. This represents the graph-level vector after attention calculation, Linear(.) represents a fully connected layer, and T represents the transpose of the matrix. This is the scaling factor in the attention mechanism.
7. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 6, characterized in that, Step 6 specifically involves: The calculated and The features are summed and then passed through multiple fully connected layers to construct a feature encoder, resulting in a feature embedding that simultaneously contains state information and information from graph-level vectors. This is the final representation vector.
8. The satellite communication fault root cause analysis method based on a combination of large and small models according to claim 7, characterized in that, Step 7 specifically includes: Feature embedding The input decoding layer is used to locate the root cause of satellite communication faults, and the root cause analysis results are obtained: in, The results of the root cause analysis should include at least one of the following: the faulty node, the cause of the fault, the associated equipment, the affected links, the fault propagation path, and the recommended objects to be investigated.