Multi-agent collaboration and rag enhanced fault diagnosis method and system for satellite communication network

Abstract

This invention discloses a method and system for fault diagnosis of satellite communication networks based on multi-agent collaboration and RAG enhancement, belonging to the fields of artificial intelligence large-scale models and satellite communication network operation and maintenance technology. This invention constructs a collaborative architecture of a task planning agent, a parsing agent, a reasoning agent, and a feedback optimization agent. Combined with a satellite communication domain-enhanced RAG mechanism, it achieves fully automated diagnosis from task input, PCAP multi-dimensional segmentation and parsing, general / satellite-specific feature extraction, protocol expert verification, RAG dual retrieval (semantic vector + keywords), to root cause fusion reasoning. The accuracy of localization is improved through multi-agent hierarchical collaboration and a divergence self-healing mechanism, and the retrieval weights and collaborative strategies are continuously optimized using reinforcement learning closed-loop feedback. This invention reduces the dependence on domain experts in operation and maintenance, and has high real-time performance, strong generalization, and traceability, and can be widely applied to the intelligent operation and maintenance of low-Earth orbit constellations, gateway stations, and terminal nodes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence large-scale models and satellite communication network operation and maintenance technology, specifically to a method and system for fault diagnosis of satellite communication networks with multi-agent collaboration and RAG enhancement, which is particularly suitable for scenarios involving the automated parsing, anomaly identification, and root cause location of massive formatted and unformatted data of systems, network devices, and terminal nodes. Background Technology

[0002] Satellite communication systems are a crucial component of integrated space-ground information and communication networks, widely serving critical areas such as emergency communications, wide-area coverage, aviation and maritime services, and access for users in remote areas. With the rapid evolution of technologies such as low-Earth orbit internet constellations and integrated space-ground networking, satellite communication systems are developing towards multi-satellite collaboration, high-speed broadband, dynamic topology, and global interconnection. Network scale continues to expand, with an increasing number of network devices, terminal nodes, and service types. Network traffic is growing exponentially, protocol adaptation is complex, and distributed operation and maintenance in different locations directly lead to problems such as multi-factor coupling, difficulty in fault location, and wide-ranging impact of system failures. This presents system operation and maintenance with core pain points including high real-time requirements, high complexity in fault location, high operational barriers, and high collaborative costs.

[0003] Currently, fault diagnosis of networks, devices, and terminal nodes mainly relies on traditional technical solutions, the core implementation methods of which include three categories: manual analysis, rule engine-driven analysis, and automated script scheduling analysis. Manual analysis refers to operations and maintenance personnel directly viewing the packet details of PCAP files using tools such as Wireshark and tcpdump. They then filter, disassemble, and statistically analyze the PCAP files, and network operations and maintenance personnel with professional knowledge analyze the protocol fields, timing relationships, and session states of the data packets based on their experience to locate the root cause of the fault. This method can only provide isolated data packet statistical results and cannot achieve multi-dimensional data correlation reasoning. It relies on the accumulated experience of operators across sessions and protocols, and has extremely poor adaptability to complex protocols and massive traffic scenarios. It is difficult to identify hidden faults and cross-layer related faults, resulting in low fault location efficiency, high false positive rate, and difficulty in meeting the real-time operation and maintenance needs of large-scale networks.

[0004] Rule-driven analysis is based on preset thresholds, signature libraries, regular expressions, and other rules to match structured data parsed from PCAP files, identifying known types of network faults (such as TCP retransmissions and DNS resolution failures) or security threats (such as port scanning and SYN flooding attacks). This method requires manual pre-definition of a fully covered set of rules and fault types, and it lacks adaptability to new protocols, proprietary protocols, and unknown faults / attacks. Furthermore, the rule maintenance cost is high, making it difficult to cope with the dynamic iteration of network protocols and the diversified evolution of fault modes.

[0005] The functionality of automated script scheduling technology is limited to a single stage (such as parsing, filtering, and alarming). It requires pre-prepared operation processes for switching between multiple tools and systems, lacks a unified task scheduling and tool collaboration framework, lacks an autonomous decision engine to support the decomposition and execution of multi-step tasks, and cannot dynamically adjust tool calling strategies based on analysis results. It has not solved the conversion barrier of "binary data → model understandable semantics".

[0006] In the traditional technical solutions listed above, PCAP files, as the standard storage format for network data packets, can completely record the raw network data from the link layer to the application layer, including key information such as packet header information (e.g., IP address, port number, protocol type, flags), data payload, timestamps, and flow identifiers. They are the core data carrier for network fault location, security threat detection, performance bottleneck analysis, and protocol compliance verification, and are widely used in the operation and maintenance and security management of various network environments. How to achieve rapid detection, accurate location, root cause identification, and self-healing recovery of system faults through parsing PCAP data packets in traditional technical solutions, and meet the system's extremely high requirements for real-time performance, accuracy, and automation in fault diagnosis technology, has become a crucial problem that urgently needs to be solved.

[0007] In recent years, artificial intelligence technologies represented by Large Language Models (LLMs) have made groundbreaking progress, possessing powerful capabilities in natural language understanding, long-context reasoning, knowledge fusion, and generalization. The rise of multi-agent collaborative architectures has further expanded the capabilities of large model technologies. By using large models as the core of intelligent agents, coupled with specialized functional agents such as parsing agents, planning agents, retrieval agents, and tool-calling agents, a collaborative and efficient technical system can be constructed. Combining large model technology with multi-agent collaborative architectures enables a technological upgrade from "human / rule-driven" to "AI-driven autonomous reasoning," breaking through the limitations of traditional solutions in data processing modes, protocol adaptation range, and anomaly detection capabilities, becoming an important technological development direction in the field of network operations and maintenance. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for fault diagnosis of satellite communication networks based on multi-agent collaboration and RAG enhancement. It integrates satellite communication domain knowledge to enhance the RAG mechanism and optimize the large model's ability to understand, plan, and decompose satellite communication domain tasks. It designs a multi-agent architecture (containing multiple specialized functional agents) and a hierarchical collaboration strategy, and constructs a technical system of "multi-agent collaboration + satellite communication domain enhanced RAG mechanism". This enables rapid parsing of massive satellite communication network PCAP data, accurate anomaly identification, and efficient fault root cause location, etc., and empowers the entire operation and maintenance process with intelligence, significantly improving the system's operation and maintenance efficiency and reliability.

[0009] The technical solution adopted in this invention is as follows: A fault diagnosis method for satellite communication networks with multi-agent collaboration and RAG enhancement includes the following steps: Step 1: Receive user natural language commands or system alarms, start the data acquisition engine, obtain auxiliary information such as the status of the satellite network link, network equipment and terminal node parameters through the standardized communication protocol adapted to satellite communication, and preliminarily determine the fault type by combining the pre-generated fault type retrieval library. The task planning agent allocates sub-task chains and schedules the subsequent required agents according to the fault type. Step 2: The task planning agent calls the parsing agent, which includes a PCAP parsing sub-agent, a feature extraction sub-agent, and a data preprocessing sub-agent. The three work in parallel and collaboratively to process the data into the standard format required for intelligent fault analysis, providing standardized input for enhanced RAG retrieval and large model fault reasoning. Step 3: The task planning agent calls the reasoning agent, which includes a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent. The reasoning agent, in turn, performs logical coordination of protocol verification, RAG case retrieval, data supplementation, and root cause fusion to analyze the standard format data and iteratively supplemented data generated in Step 2 and perform fault root cause reasoning to generate fault root causes and RAG retrieval basis. Step 4: The task planning agent calls the report generation agent and the feedback optimization agent to synchronize the root cause of the fault and the RAG retrieval basis to the report generation agent and the feedback optimization agent, and coordinates them according to the logic of knowledge base update after report generation. Step 5: Optimize the agent through feedback, improve fault location accuracy, RAG knowledge base, RAG retrieval matching degree and agent collaboration strategy to adapt to the dynamic changes of the satellite communication network.

[0010] Furthermore, the process of step 1 is as follows: Step 11: Start the information acquisition engine and generate a unified format acquisition request packet for multiple acquisition objects based on the adapted satellite communication standardization protocol; Step 12: Perform dual verification of completeness and validity on the structured auxiliary information dataset output in Step 11 based on the preset data verification rule base, and adopt an automatic completion mechanism for missing or invalid data. Step 13: Based on the standardized auxiliary information dataset output in Step 12, call the pre-generated fault type retrieval library to preliminarily determine the fault scenario type; Step 14: Based on the fault scenario type determined in Step 13, and combined with the influencing factors corresponding to the task triggering method, a weighted scoring method is used to calculate the priority score, which is divided into three levels: high, medium, and low, for subsequent system resource allocation and task scheduling; the task triggering methods include monitoring alarms and user commands.

[0011] Furthermore, the process of step 2 is as follows: Step 21: Employ a multi-dimensional segmentation strategy to adapt to the model context window. The segmentation strategy includes: a) Session / stream-based segmentation: Split the large PCAP file by 5-tuples or bidirectional sessions to ensure that each generated CSV represents a complete session interaction; b) Protocol-layer-based segmentation: Filter and extract specific protocols for satellite 5G base stations, including RRC, NAS, and NGAP; c) Time window and overlapping context-based segmentation: Segment by time period and introduce a sliding window strategy to retain overlapping content between consecutive blocks. Step 22: Call the PCAP parsing sub-agent. The PCAP parsing sub-agent contains a satellite communication-adapted parsing tool, which is used to read GB / TB-level satellite communication PCAP files. Based on the built-in satellite communication private protocol parsing rule library, it parses satellite communication-specific fields with satellite-to-ground link identifier, beam number and modulation and coding scheme. At the same time, it extracts general fields with five-tuple, data packet size and transmission timing, and outputs structured protocol parsing results. Step 23: Based on the analysis results, call the feature extraction sub-agent to extract general features and satellite communication-specific features, and generate a standardized feature matrix and RAG retrieval vector; the general features include TCP retransmission rate and UDP packet loss rate; the satellite communication-specific features include link switching frequency, bit error rate, latency jitter and ephemeris parameter correlation values; the RAG retrieval vector contains fault feature labels and protocol type identifiers; Step 24: Invoke the data preprocessing sub-agent to perform noise filtering and missing value completion preprocessing. At the same time, dynamically adjust the feature weights according to the satellite communication network status, and finally output the structured traffic summary, feature vector and RAG retrieval vector.

[0012] Furthermore, the process of step 3 is as follows: Step 31: The task planning agent breaks down the core task into a sub-task chain of protocol legality verification, feature association analysis, RAG retrieval, and root cause localization, and assigns it to the corresponding agent. At the same time, it generates a satellite communication scenario-based RAG retrieval instruction and clarifies the retrieval priority. The retrieval priority is set in the order of on-board fault > link fault > ground station fault. Step 32: Based on the constructed RAG retrieval protocol library containing satellite communication standards and private protocol specifications, call the protocol expert sub-agent to verify the legality of protocol fields and the rationality of timing, identify protocol anomalies and output anomaly types and related features, and synchronize them to the fault diagnosis sub-agent; Step 33: Invoke the case retrieval sub-agent to perform RAG typical case matching retrieval, specifically including: a. Knowledge Base Access: Access the "Satellite Communication Fault Case Library + Protocol Knowledge Base". The satellite communication fault cases in the case library all contain associated feature vectors, root cause analysis and handling solutions. The protocol knowledge base integrates satellite communication network standards and private protocol documents. b. Retrieval Layer: A dual retrieval method combining semantic vector retrieval and keyword exact matching is adopted. Semantic vector retrieval is used to match RAG retrieval vectors with case feature vectors, while keyword exact matching is used to match beam numbers and fault phenomena. c. Output Results: Sort the search results by priority of satellite communication fault handling, and output similar cases and handling solutions for reference. The similar cases include case ID and matching degree. Step 34: Integrate the feature data generated in Step 2, the abnormal results of the protocol expert sub-agent, and the RAG retrieval results of the case retrieval sub-agent, call the fault diagnosis sub-agent, infer the candidate root causes of the fault through the multi-dimensional reasoning model, and output the root cause type and confidence level. Step 35: If there is a disagreement among agents regarding the root cause, the case retrieval sub-agent expands the RAG retrieval case sample, the data acquisition engine supplements the system operation data, and the reasoning and retrieval verification are re-executed until the confidence level is not less than the preset threshold; the disagreement includes the situation where the protocol expert sub-agent determines "protocol failure" while the fault diagnosis sub-agent determines "link failure".

[0013] Furthermore, the process of step 4 is as follows: Step 41: Call the report generation agent to automatically generate a structured diagnostic report based on the reasoning results and RAG search criteria. Highlight the priority of satellite communication fault handling and emergency procedures, and clearly mark similar cases associated with RAG search to ensure the traceability of results. Step 42: Record task allocation records, retrieval logs, and divergence handling results during the agent collaboration process. Based on the overall fault handling effect, call the feedback optimization agent to label RAG case matching degree and optimize RAG retrieval weight; the feedback optimization agent includes a reinforcement learning algorithm. Step 43: The diagnostic report is pushed to the user or operation and maintenance system, and the optimized search weights, new fault scenarios and rules are synchronized to the RAG knowledge base.

[0014] Furthermore, the process of step 5 is as follows: Step 51: Based on the collaborative process data, adjust the sub-task chain splitting rules, feature weight adjustment thresholds, and RAG retrieval parameters to optimize the agent collaborative strategy. Step 52: For satellite communication application scenarios that are added from time to time, dynamically update the intelligent agent adaptation rules and RAG retrieval strategy.

[0015] Furthermore, the fault type database includes key dimensions such as on-board faults, link faults, ground station faults, and equipment and terminal node faults; the RAG search criteria include similar case numbers.

[0016] Furthermore, the task planning agent undertakes the core scheduling and decision-making functions of the multi-agent collaborative architecture; the parsing agent focuses on the efficient parsing and format standardization of PCAP binary data; and the reasoning agent focuses on cross-layer and temporal dimension correlation analysis and rapid matching of similar cases in the fault knowledge base.

[0017] A multi-agent collaborative and RAG-enhanced satellite communication network fault diagnosis system, used to implement a multi-agent collaborative and RAG-enhanced satellite communication network fault diagnosis method, characterized by comprising: The task planning agent is used to receive user natural language commands or system alarms, start the data acquisition engine, make a preliminary judgment on the fault type, and perform sub-task chain allocation and agent scheduling. The parsing agent consists of a PCAP parsing sub-agent, a feature extraction sub-agent, and a data preprocessing sub-agent. These three agents work in parallel and in collaboration to process data into the standard format required for intelligent fault analysis. The reasoning agent, comprising a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent, is used for intelligent analysis and fault diagnosis based on the logic of protocol verification, RAG case retrieval, data supplementation, and root cause fusion. Report generation agent and feedback optimization agent are used to generate structured diagnostic reports, update the knowledge base, and perform iterative optimization; The data acquisition engine is used to acquire auxiliary information from the satellite communication network through standardized communication protocols adapted to satellite communication. An enhanced RAG mechanism for the satellite communication domain, including a satellite communication fault case library and a protocol knowledge base, is used to support typical case matching and retrieval by the case retrieval sub-agent; The task planning agent is communicatively connected to the parsing agent, the reasoning agent, the report generating agent, and the feedback optimization agent.

[0018] Furthermore, the PCAP parsing sub-agent includes a satellite communication-adapted parsing tool with a built-in satellite communication private protocol parsing rule library, used to parse satellite-to-ground link identifier, beam number, modulation and coding scheme satellite communication exclusive fields, as well as five-tuple, data packet size, and transmission timing general fields; The feature extraction sub-agent is used to extract general features and satellite communication-specific features, and generate a standardized feature matrix and RAG retrieval vector; the general features include TCP retransmission rate and UDP packet loss rate; the satellite communication-specific features include link switching frequency, bit error rate, latency jitter, and ephemeris parameter correlation values; The data preprocessing sub-agent is used to perform noise filtering, missing value completion, and dynamically adjust feature weights according to the satellite communication network status, and finally outputs structured traffic summaries, feature vectors, and RAG retrieval vectors.

[0019] Furthermore, the case retrieval sub-agent is used to perform a dual retrieval combining semantic vector retrieval and keyword precise matching, and outputs similar cases and handling solutions for reference according to the priority of satellite communication fault handling. The feedback optimization agent employs a reinforcement learning algorithm to automatically label RAG case matching degrees and optimize RAG retrieval weights, as well as synchronize the optimized retrieval weights, new fault scenarios, and rules to the RAG knowledge base.

[0020] Furthermore, it also includes a multi-dimensional segmentation module for performing session / stream-based segmentation, protocol-layer-based segmentation, and time-window and overlapping context-based segmentation on PCAP files to accommodate large model context windows.

[0021] The advantages of this invention compared to the prior art are as follows: This invention significantly improves parsing efficiency: by combining multi-agent parallel collaboration with RAG precise retrieval, only key fault-related data is extracted, which can reduce the parsing time of satellite communication PCAP data from hours to minutes, thus greatly improving efficiency; This invention has strong adaptability to satellite communication scenarios: through a dedicated satellite communication protocol library and retrieval strategy, it can achieve full coverage of fault scenarios for dedicated protocols and typical systems, devices and terminal nodes; This invention offers high fault location accuracy: the RAG enhancement mechanism enables multi-dimensional reasoning through "feature matching + case verification," improving the accuracy of fault location specific to satellite communications, and the retrieval basis is traceable, solving the problem of "black box" reasoning. This invention features a high degree of automation throughout the entire process: from PCAP parsing to report generation, no manual intervention is required. It supports natural language interaction and interface with satellite communication devices, reducing reliance on domain experts and significantly reducing the workload of manual data processing and analysis. It has broad engineering application prospects and promotional value. This invention has excellent dynamic scalability: through the RAG feedback iteration mechanism, the retrieval weights are adjusted in real time to adapt to the dynamic changes of the satellite communication network, eliminating the need for frequent maintenance of the knowledge base and reducing operation and maintenance costs. Attached Figure Description

[0022] The accompanying drawings, which are provided to further illustrate the invention and constitute a part of this invention, are not intended to limit the invention. In the drawings: Figure 1 This is a flowchart of an intelligent fault diagnosis method for satellite communication networks based on multi-agent collaboration and RAG enhancement proposed in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0024] The core of the technical solution of this invention is a "multi-agent collaborative architecture + satellite communication domain enhanced RAG mechanism". It constructs a task planning agent that serves as the core scheduling and decision-making center of the multi-agent collaborative architecture, a parsing agent that focuses on efficient parsing and format standardization of PCAP binary data, and a reasoning agent that focuses on cross-layer and time-series correlation analysis and rapid matching of similar cases in the fault knowledge base. Based on the satellite communication network workflow, the invention designs agent task allocation strategies, cross-layer interaction rules and conflict coordination mechanisms to achieve efficient linkage of multiple agents. Combined with the satellite communication exclusive knowledge base and set enhanced retrieval strategies, it realizes the technical upgrade of PCAP parsing and fault diagnosis from "manual / rule-driven" to "AI autonomous reasoning driven".

[0025] Example 1: Reference Figure 1 As shown, this invention provides a method for fault diagnosis of satellite communication networks based on multi-agent collaboration and RAG enhancement, comprising the following steps: Step 1: Task Input. Upon receiving user natural language commands or system alarms, the data acquisition engine is activated. Auxiliary information such as satellite communication network link status and network parameters is obtained through standardized communication protocols adapted to satellite communication. Combined with a pre-generated fault type retrieval library (including key dimensions such as onboard faults, link faults, and ground station faults), the fault type is initially determined. The task planning agent allocates sub-task chains and schedules subsequent required agents based on the fault type.

[0026] Step 2: Data Preprocessing and Feature Extraction. The task planning agent calls the parsing agent, which includes the PCAP parsing sub-agent, the feature extraction sub-agent, and the data preprocessing sub-agent. These three agents work in parallel and collaboratively to process the data into the standard format required for intelligent fault analysis, providing standardized input for enhanced RAG retrieval and large-scale model fault reasoning.

[0027] Step 3: Intelligent Analysis and Fault Diagnosis. The task planning agent calls upon the reasoning agent, which includes a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent. Following the logical collaboration of "protocol verification → RAG case retrieval → data supplementation → root cause fusion," the agent analyzes the standard format data generated in Step 2 and the iteratively supplemented data in Step 3, and performs fault root cause reasoning to generate fault root causes and RAG retrieval criteria (similar case numbers, etc.).

[0028] Step 4: Collaborative Output. The task planning agent calls the report generation agent and the feedback optimization agent, and integrates the root cause of the fault and the RAG retrieval basis into the agents in this process, collaborating according to the logic of "report generation → knowledge base update". Step 5: Iterative Optimization. The agent is optimized through feedback, improving fault location accuracy, RAG knowledge base, RAG retrieval matching degree, and agent collaboration strategy to adapt to the dynamic changes in the satellite communication network.

[0029] Step 1 specifically includes the following sub-steps: Step 11: Auxiliary Information Acquisition Initialization. Start the information acquisition engine and generate a unified format acquisition request packet for multiple acquisition objects in the system based on the adapted satellite communication standardization protocol; Step 12: Data collection verification and completion. Based on the preset "data verification rule base", the structured auxiliary information dataset output in Step 11 is subjected to dual verification of completeness and validity. An automatic completion mechanism is used for missing or invalid data to ensure that the data meets the requirements of subsequent fault classification and priority calculation. Step 13: Preliminary Classification of Satellite Communication Fault Scenarios. Based on the standardized auxiliary information dataset output in Step 12, the pre-generated "fault type retrieval library" is invoked. Using the matching and judgment logic of "core conditions + auxiliary conditions + pre-defined fault judgment priorities", the fault scenario types are preliminarily determined, providing a basis for subsequent priority calculation. Step 14: Task Priority Calculation and Sorting. Based on the fault scenario type determined in Step 13, and combined with the influencing factors corresponding to the task triggering method (monitoring alarm / user command), a weighted scoring method is used to calculate the priority score, which is divided into three levels: high, medium, and low, for subsequent system resource allocation and task scheduling.

[0030] Step 2 specifically includes the following sub-steps: Step 21: File Segmentation. The system employs a multi-dimensional segmentation strategy to adapt to the model context window; (1) Session / stream segmentation based: Principle: Use tools to segment large PCAPs by 5-tuples or bidirectional sessions.

[0031] Advantages: Ensures that each generated CSV represents a complete session interaction, maintains consistency of fault context, and avoids information fragmentation due to cross-block issues.

[0032] (2) Protocol layer-based segmentation: Principle: Filter and extract specific protocols (such as RRC, NAS, NGAP) for satellite 5G base stations.

[0033] Application: If fault location focuses on a specific protocol (such as RRC chain establishment failure), only relevant layer data is extracted to reduce noise.

[0034] (3) Based on time window and overlapping context: Principle: Divide the data into time segments (e.g., every 60 seconds) and introduce a "sliding window" strategy.

[0035] Implementation: Preserve overlapping content between consecutive blocks (such as a later block containing the tail of the previous block) to ensure that LLM maintains a coherent contextual understanding when analyzing cross-block transactions.

[0036] Step 22: Call the PCAP parsing sub-agent, which includes a satellite communication-adapted parsing tool (such as the Wireshark satellite communication plugin), which can read GB / TB level satellite communication PCAP files. Based on the built-in satellite communication private protocol parsing rule library, it parses satellite communication-specific fields such as satellite-to-ground link identifier, beam number, modulation and coding scheme, while extracting common fields such as quintuple, data packet size, and transmission timing, and outputs structured protocol parsing results. Step 23: Based on the parsing results, call the feature extraction sub-agent to extract general features (TCP retransmission rate, UDP packet loss rate) and satellite communication-specific features (link switching frequency, bit error rate, latency jitter, ephemeris parameter correlation value) to generate a standardized feature matrix and RAG retrieval vector (including fault feature labels and protocol type identifiers). Step 24: Call the data preprocessing sub-agent to perform preprocessing processes such as noise filtering and missing value completion. At the same time, dynamically adjust the feature weights according to the satellite communication network status, and finally output the structured traffic summary, feature vector and RAG retrieval vector. Furthermore, step 3 specifically includes the following sub-steps: Step 31: The task planning agent breaks down the core task into a sub-task chain of "protocol legality verification → feature association analysis → RAG retrieval → root cause localization", assigns it to the corresponding agent, and generates satellite communication scenario-based RAG retrieval instructions to clarify the retrieval priority (e.g., on-board faults > link faults > ground station faults). Step 32: Based on the constructed RAG retrieval protocol library containing satellite communication standards and private protocol specifications, call the protocol expert sub-agent to verify the legality of protocol fields and the rationality of timing, identify protocol anomalies and output the anomaly type and related features, and synchronize to the fault diagnosis sub-agent; Step 33: Invoke the case retrieval sub-agent to perform RAG typical case matching retrieval; (1) Knowledge base access: Access the "Satellite Communication Fault Case Library + Protocol Knowledge Base". The satellite communication fault cases in the case library all contain associated feature vectors, root cause analysis and handling solutions. The protocol knowledge base integrates satellite communication network standards and private protocol documents. (2) Retrieval layer: semantic vector retrieval (matching RAG retrieval vectors with case feature vectors) + keyword precise matching (beam number, fault phenomenon, etc.); (3) Results output: Sort the search results according to the priority of satellite communication fault handling, and output similar cases (including case ID and matching degree) and handling plan references; Step 34: Integrate the feature data generated in Step 2, the abnormal results of the protocol expert sub-agent, and the RAG retrieval results of the case retrieval sub-agent, call the fault diagnosis sub-agent, infer the candidate root causes of the fault through the multi-dimensional reasoning model, and output the root cause type and confidence level. Step 35: If there is a disagreement among agents regarding the root cause (e.g., the protocol expert determines "protocol failure" and the fault diagnosis sub-agent determines "link failure"), the case retrieval sub-agent expands the RAG retrieval case sample, the data acquisition engine supplements the system operation data, and the reasoning and retrieval verification are re-executed until the confidence level is not less than the preset threshold. Furthermore, step 4 specifically includes the following sub-steps: Step 41: Call the report generation agent to automatically generate a structured diagnostic report based on the reasoning results and RAG search criteria, highlighting the priority of satellite communication fault handling and emergency procedures, clearly marking similar cases associated with RAG search, and ensuring the traceability of results; Step 42: Record data such as task allocation records, retrieval logs, and divergence handling results during the agent collaboration process. Based on the comprehensive fault handling effect, call feedback optimization agents (including reinforcement learning and other algorithms) to automatically label RAG case matching degree and optimize RAG retrieval weight. Step 43: The diagnostic report is pushed to the user / operations system, and the optimized search weights, new fault scenarios and rules are synchronized to the RAG knowledge base. Furthermore, step 5 specifically includes the following sub-steps: Step 51: Based on the collaborative process data, adjust the sub-task chain splitting rules, feature weight adjustment thresholds, RAG retrieval parameters, etc., to optimize the agent collaborative strategy; Step 52: For satellite communication application scenarios that are added from time to time, dynamically update the intelligent agent adaptation rules and RAG retrieval strategy.

[0037] Example 2: The satellite communication network fault diagnosis system based on multi-agent collaboration and RAG enhancement in this example includes: The task planning agent is used to receive user natural language commands or system alarms, start the data acquisition engine, make a preliminary judgment on the fault type, and perform sub-task chain allocation and agent scheduling. The parsing agent consists of a PCAP parsing sub-agent, a feature extraction sub-agent, and a data preprocessing sub-agent. These three agents work in parallel and in collaboration to process data into the standard format required for intelligent fault analysis. The reasoning agent includes a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent, which is used to perform intelligent analysis and fault diagnosis according to the logic of protocol verification → RAG case retrieval → data supplementation → root cause fusion. Report generation agent and feedback optimization agent are used to generate structured diagnostic reports, update the knowledge base, and perform iterative optimization; The data acquisition engine is used to acquire auxiliary information from the satellite communication network through standardized communication protocols adapted to satellite communication. An enhanced RAG mechanism for the satellite communication domain, including a satellite communication fault case library and a protocol knowledge base, is used to support typical case matching and retrieval by the case retrieval sub-agent; The multidimensional segmentation module is used to perform session / stream-based segmentation, protocol-layer-based segmentation, and time-window and overlapping context-based segmentation on PCAP files to accommodate large model context windows.

[0038] The task planning agent is communicatively connected to the parsing agent, the reasoning agent, the report generating agent, and the feedback optimization agent.

[0039] The PCAP parsing sub-agent includes a satellite communication-adapted parsing tool with a built-in satellite communication private protocol parsing rule library, used to parse satellite-to-ground link identifier, beam number, modulation and coding scheme satellite-specific fields, as well as five-tuple, data packet size, and transmission timing general fields; The feature extraction sub-agent is used to extract general features and satellite communication-specific features, and generate a standardized feature matrix and RAG retrieval vector; the general features include TCP retransmission rate and UDP packet loss rate; the satellite communication-specific features include link switching frequency, bit error rate, latency jitter, and ephemeris parameter correlation values; The data preprocessing sub-agent is used to perform noise filtering, missing value completion, and dynamically adjust feature weights according to the satellite communication network status, and finally outputs structured traffic summaries, feature vectors, and RAG retrieval vectors.

[0040] The case retrieval sub-agent is used to perform a dual retrieval combining semantic vector retrieval and keyword precise matching, and outputs similar cases and handling solutions for reference according to the priority of satellite communication fault handling. The feedback optimization agent employs a reinforcement learning algorithm to automatically label RAG case matching degrees and optimize RAG retrieval weights, as well as synchronize the optimized retrieval weights, new fault scenarios, and rules to the RAG knowledge base.

[0041] This invention combines large model technology, RAG enhanced retrieval, and multi-agent collaborative architecture to automate the entire fault diagnosis process in satellite communication-specific scenarios. It breaks through the limitations of traditional "manual / rule-driven" methods in terms of data processing scale, protocol adaptation range, anomaly detection capability, and contextual reasoning, and becomes an important technological development direction for intelligent network operation and maintenance of satellite communication systems.

Claims

1. A fault diagnosis method for satellite communication networks based on multi-agent collaboration and RAG enhancement, characterized in that, Specifically, the following steps are included: Step 1: Receive user natural language commands or system alarms, start the data acquisition engine, obtain auxiliary information such as the status of the satellite network link, network equipment and terminal node parameters through the standardized communication protocol adapted to satellite communication, and preliminarily determine the fault type by combining the pre-generated fault type retrieval library. The task planning agent allocates sub-task chains and schedules the subsequent required agents according to the fault type. Step 2: The task planning agent calls the parsing agent, which includes a PCAP parsing sub-agent, a feature extraction sub-agent, and a data preprocessing sub-agent. The three work in parallel and collaboratively to process the data into the standard format required for intelligent fault analysis, providing standardized input for enhanced RAG retrieval and large model fault reasoning. Step 3: The task planning agent calls the reasoning agent, which includes a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent. The reasoning agent, in turn, performs logical coordination of protocol verification, RAG case retrieval, data supplementation, and root cause fusion to analyze the standard format data and iteratively supplemented data generated in Step 2 and perform fault root cause reasoning to generate fault root causes and RAG retrieval basis. Step 4: The task planning agent calls the report generation agent and the feedback optimization agent to synchronize the root cause of the fault and the RAG retrieval basis to the report generation agent and the feedback optimization agent, and coordinates them according to the logic of knowledge base update after report generation. Step 5: Optimize the agent through feedback, improve fault location accuracy, RAG knowledge base, RAG retrieval matching degree and agent collaboration strategy to adapt to the dynamic changes of the satellite communication network.

2. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The process of step 1 is as follows: Step 11: Start the information acquisition engine and generate a unified format acquisition request packet for multiple acquisition objects based on the adapted satellite communication standardization protocol; Step 12: Perform dual verification of completeness and validity on the structured auxiliary information dataset output in Step 11 based on the preset data verification rule base, and adopt an automatic completion mechanism for missing or invalid data. Step 13: Based on the standardized auxiliary information dataset output in Step 12, call the pre-generated fault type retrieval library to preliminarily determine the fault scenario type; Step 14: Based on the fault scenario type determined in Step 13, and combined with the influencing factors corresponding to the task triggering method, a weighted scoring method is used to calculate the priority score, which is divided into three levels: high, medium, and low, for subsequent system resource allocation and task scheduling; the task triggering methods include monitoring alarms and user commands.

3. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The process of step 2 is as follows: Step 21: Employ a multi-dimensional segmentation strategy to adapt to the model context window. The segmentation strategy includes: a) Session / stream-based segmentation: Split the large PCAP file by 5-tuples or bidirectional sessions to ensure that each generated CSV represents a complete session interaction; b) Protocol-layer-based segmentation: Filter and extract specific protocols for satellite 5G base stations, including RRC, NAS, and NGAP; c) Time window and overlapping context-based segmentation: Segment by time period and introduce a sliding window strategy to retain overlapping content between consecutive blocks. Step 22: Call the PCAP parsing sub-agent. The PCAP parsing sub-agent contains a satellite communication-adapted parsing tool, which is used to read GB / TB-level satellite communication PCAP files. Based on the built-in satellite communication private protocol parsing rule library, it parses satellite communication-specific fields with satellite-to-ground link identifier, beam number and modulation and coding scheme. At the same time, it extracts general fields with five-tuple, data packet size and transmission timing, and outputs structured protocol parsing results. Step 23: Based on the analysis results, call the feature extraction sub-agent to extract general features and satellite communication-specific features, and generate a standardized feature matrix and RAG retrieval vector; the general features include TCP retransmission rate and UDP packet loss rate; the satellite communication-specific features include link switching frequency, bit error rate, latency jitter and ephemeris parameter correlation values; the RAG retrieval vector contains fault feature labels and protocol type identifiers; Step 24: Invoke the data preprocessing sub-agent to perform noise filtering and missing value completion preprocessing. At the same time, dynamically adjust the feature weights according to the satellite communication network status, and finally output the structured traffic summary, feature vector and RAG retrieval vector.

4. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 31: The task planning agent breaks down the core task into a sub-task chain of protocol legality verification, feature association analysis, RAG retrieval and root cause localization, and assigns it to the corresponding agent. At the same time, it generates a satellite communication scenario-based RAG retrieval instruction and clarifies the retrieval priority. The retrieval priority is set in the order of on-board fault > link fault > ground station fault. Step 32: Based on the constructed RAG retrieval protocol library containing satellite communication standards and private protocol specifications, call the protocol expert sub-agent to verify the legality of protocol fields and the rationality of timing, identify protocol anomalies and output anomaly types and related features, and synchronize them to the fault diagnosis sub-agent; Step 33: Invoke the case retrieval sub-agent to perform RAG typical case matching retrieval, specifically including: a. Knowledge Base Access: Access the "Satellite Communication Fault Case Library + Protocol Knowledge Base". The satellite communication fault cases in the case library all contain associated feature vectors, root cause analysis and handling solutions. The protocol knowledge base integrates satellite communication network standards and private protocol documents. b. Retrieval Layer: A dual retrieval method combining semantic vector retrieval and keyword exact matching is adopted. Semantic vector retrieval is used to match RAG retrieval vectors with case feature vectors, while keyword exact matching is used to match beam numbers and fault phenomena. c. Output Results: Sort the search results by priority of satellite communication fault handling, and output similar cases and handling solutions for reference. The similar cases include case ID and matching degree. Step 34: Integrate the feature data generated in Step 2, the abnormal results of the protocol expert sub-agent, and the RAG retrieval results of the case retrieval sub-agent, call the fault diagnosis sub-agent, infer the candidate root causes of the fault through the multi-dimensional reasoning model, and output the root cause type and confidence level. Step 35: If there is a disagreement among agents regarding the root cause, the case retrieval sub-agent expands the RAG retrieval case sample, the data acquisition engine supplements the system operation data, and the reasoning and retrieval verification are re-executed until the confidence level is not less than the preset threshold. The disagreement includes the situation where the protocol expert sub-agent determines "protocol failure" while the fault diagnosis sub-agent determines "link failure".

5. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The process of step 4 is as follows: Step 41: Call the report generation agent to automatically generate a structured diagnostic report based on the reasoning results and RAG search criteria. Highlight the priority of satellite communication fault handling and emergency procedures, and clearly mark similar cases associated with RAG search to ensure the traceability of results. Step 42: Record task allocation records, retrieval logs, and divergence handling results during the agent collaboration process. Based on the overall fault handling effect, call the feedback optimization agent to label RAG case matching degree and optimize RAG retrieval weight; the feedback optimization agent includes a reinforcement learning algorithm. Step 43: The diagnostic report is pushed to the user or operation and maintenance system, and the optimized search weights, new fault scenarios and rules are synchronized to the RAG knowledge base.

6. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The process of step 5 is as follows: Step 51: Based on the collaborative process data, adjust the sub-task chain splitting rules, feature weight adjustment thresholds, and RAG retrieval parameters to optimize the agent collaborative strategy. Step 52: For satellite communication application scenarios that are added from time to time, dynamically update the intelligent agent adaptation rules and RAG retrieval strategy.

7. The satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement according to claim 1, characterized in that, The fault type database includes key dimensions such as on-board faults, link faults, ground station faults, and equipment and terminal node faults; the RAG search criteria include similar case numbers.

8. The method for fault diagnosis of satellite communication networks based on multi-agent cooperation and RAG enhancement according to claim 1, characterized in that, The task planning agent undertakes the core scheduling and decision-making functions of the multi-agent collaborative architecture; the parsing agent focuses on the efficient parsing and format standardization of PCAP binary data; and the reasoning agent focuses on cross-layer and temporal dimension correlation analysis and rapid matching of similar cases in the fault knowledge base.

9. A satellite communication network fault diagnosis system based on multi-agent collaboration and RAG enhancement, used to implement the satellite communication network fault diagnosis method based on multi-agent collaboration and RAG enhancement as described in any one of claims 1 to 8, characterized in that, include: The task planning agent is used to receive user natural language commands or system alarms, start the data acquisition engine, make a preliminary judgment on the fault type, and perform sub-task chain allocation and agent scheduling. The parsing agent consists of a PCAP parsing sub-agent, a feature extraction sub-agent, and a data preprocessing sub-agent. These three agents work in parallel and in collaboration to process data into the standard format required for intelligent fault analysis. The reasoning agent, comprising a protocol expert sub-agent, a case retrieval sub-agent, and a fault diagnosis sub-agent, is used for intelligent analysis and fault diagnosis based on the logic of protocol verification, RAG case retrieval, data supplementation, and root cause fusion. Report generation agent and feedback optimization agent are used to generate structured diagnostic reports, update the knowledge base, and perform iterative optimization; The data acquisition engine is used to acquire auxiliary information from the satellite communication network through standardized communication protocols adapted to satellite communication. An enhanced RAG mechanism for the satellite communication domain, including a satellite communication fault case library and a protocol knowledge base, is used to support typical case matching and retrieval by the case retrieval sub-agent; The task planning agent is communicatively connected to the parsing agent, the reasoning agent, the report generating agent, and the feedback optimization agent.

10. The satellite communication network fault diagnosis system based on multi-agent collaboration and RAG enhancement according to claim 9, characterized in that, The PCAP parsing sub-agent includes a satellite communication-adapted parsing tool with a built-in satellite communication private protocol parsing rule library, used to parse satellite-to-ground link identifier, beam number, modulation and coding scheme satellite-specific fields, as well as five-tuple, data packet size, and transmission timing general fields; The feature extraction sub-agent is used to extract general features and satellite communication-specific features, and generate a standardized feature matrix and RAG retrieval vector; the general features include TCP retransmission rate and UDP packet loss rate; the satellite communication-specific features include link switching frequency, bit error rate, latency jitter, and ephemeris parameter correlation values; The data preprocessing sub-agent is used to perform noise filtering, missing value completion, and dynamically adjust feature weights according to the satellite communication network status, and finally outputs structured traffic summaries, feature vectors, and RAG retrieval vectors.

Images