An end-edge-cloud collaborative satellite network autonomous operation and maintenance architecture and method

Abstract

The application discloses a satellite network autonomous operation and maintenance architecture and method facing end-edge-cloud cooperation, and belongs to the technical field of satellite network operation and maintenance; the operation and maintenance method comprises the following steps: acquiring fault information in a satellite network; performing semantic coding on the fault information to obtain a fault semantic feature vector; inputting the fault semantic feature vector into an intelligent hierarchical decision module, calculating corresponding expected benefits of fault tasks in an end layer, an edge layer and a cloud layer, and determining a target disposal hierarchy; calling a corresponding disposal engine according to the target disposal hierarchy to complete fault cooperative disposal; evaluating a fault disposal result, performing cross-hierarchy autonomous upgrading when the target disposal hierarchy cannot complete a fault disposal task; calculating a feedback reward according to a fault recovery result, updating an intelligent hierarchical decision model online, and realizing continuous optimization of a hierarchical decision strategy. The application can realize autonomous hierarchical decision and cross-hierarchy cooperative disposal of fault tasks, and effectively improves the autonomous operation and maintenance capability of a large-scale satellite network.

Description

Technical Field

[0001] This invention belongs to the field of satellite network operation and maintenance technology, specifically relating to a satellite network autonomous operation and maintenance architecture and method oriented towards end-edge-cloud collaboration. Background Technology

[0002] Large-scale satellite networks, as a crucial infrastructure for future integrated space-air-ground information networks, have been widely applied in satellite internet, converged space-ground communications, global IoT, and space information services. With the continuous expansion of low-Earth orbit satellite constellations, rapid and dynamic changes in network topology, and the emergence of new services, the operating environment of satellite networks is becoming increasingly complex, placing higher demands on network reliability and service continuity. Against this backdrop, achieving autonomous operation, autonomous fault recovery, and autonomous operation and maintenance management of satellite networks has become a key issue for future satellite network development. When a satellite node fails, if fault location and recovery cannot be completed in a timely manner, it will not only affect the operational status of the individual node but may also lead to link interruptions, service migration failures, and local network performance degradation, thereby impacting the service quality of the entire satellite network.

[0003] Traditional satellite networks primarily employ a centralized ground-based operation and maintenance (O&M) model. After a satellite detects a fault, it transmits the fault information back to a ground control center, where fault analysis, recovery decisions, and control command issuance are handled centrally. While this model facilitates centralized management, it is limited by the satellite-to-ground visibility window, communication bandwidth, and the centralized ground computing architecture, making it difficult to respond promptly to continuous fault events in large-scale satellite networks. Furthermore, this model heavily relies on manual analysis and fixed rules, lacking the ability for autonomous fault decision-making, recovery, and closed-loop optimization, thus failing to meet the future needs of autonomous O&M management for satellite networks. To address the high response latency, low resource utilization, and insufficient autonomy associated with centralized O&M, researchers have recently begun introducing an edge-cloud collaborative architecture into satellite network O&M. By deploying some fault handling capabilities to on-orbit satellite nodes and edge management nodes, fault handling capabilities are pushed down to the network edge, reducing reliance on the ground center and improving fault response speed. Therefore, how to fully utilize the resource advantages of heterogeneous nodes at the end, edge, and cloud layers to construct a complete autonomous O&M management closed loop has become an important research direction in the field of intelligent satellite network O&M.

[0004] Existing research mainly focuses on two areas: the construction of satellite health management systems and anomaly detection algorithms, achieving some progress in improving satellite autonomous operation capabilities and anomaly detection performance. However, the aforementioned research primarily concentrates on the health management or anomaly detection phases, and lacks systematic solutions for key autonomous operation and maintenance management issues such as autonomous hierarchical decision-making, cross-level collaborative handling, and online feedback optimization for faulty tasks under an edge-cloud collaborative architecture.

[0005] In existing technical solutions, the hierarchical distributed satellite autonomous health management method divides health management functions into subsystem layer, satellite service layer, and whole-satellite layer, enabling the migration of some health management capabilities to the satellite and improving its autonomous operation capability. However, the hierarchical division of this method is mainly designed based on the internal functional structure of the satellite, and the responsibilities of each layer are basically fixed after system deployment. It does not fully consider the differences in computing power, communication resources, and processing timeliness among end-layer, edge-layer, and cloud-layer nodes. It cannot autonomously select the target handling layer based on fault characteristics and network status, and it also lacks the ability to continuously optimize hierarchical decision-making strategies based on historical operation and maintenance feedback. As a result, it is difficult to meet the needs of dynamic decision-making and continuous evolution in the autonomous operation and maintenance management of large-scale satellite networks.

[0006] Furthermore, existing satellite anomaly detection methods based on causal structure learning or data-driven approaches can achieve anomaly detection, severity assessment, and anomaly hierarchical localization, providing effective support for fault analysis. However, these methods mainly focus on the fault detection and localization stages, lacking a unified mechanism for key issues such as how to conduct autonomous hierarchical decision-making, cross-level collaborative handling, escalation after handling failures, and historical experience feedback optimization after anomaly identification. A closed loop of autonomous operation and maintenance management covering the entire process of fault perception, hierarchical decision-making, collaborative handling, and feedback learning has not yet been formed. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention provides an architecture and method for autonomous operation and maintenance of satellite networks oriented towards edge-cloud collaboration. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for autonomous operation and maintenance of satellite networks oriented towards end-edge-cloud collaboration, comprising: The system acquires fault information generated during the operation of the satellite network, performs textual processing on the fault information, and generates natural language fault descriptions. Normalized fault semantic feature vectors are extracted from the natural language fault description using a pre-trained text encoding model. The fault semantic feature vector is processed by a feedforward neural network to estimate the expected benefits of fault handling at each level in the satellite network and generate the expected benefit value corresponding to each level; the levels include the end layer, the edge layer and the cloud layer. Based on the expected returns for each level, through The strategy is to select the target handling level for the fault task corresponding to the current fault information; The fault task is sent to the corresponding node of the target handling level, the fault handling engine of the target handling level is invoked to perform fault handling, and the corresponding first processing result is output. The first processing result is evaluated according to the preset recovery judgment rules. When the evaluation indicates that the fault recovery has failed, a hierarchical escalation mechanism is triggered to select the upgraded handling level for fault task handling until the fault recovery is successful or the highest handling level is reached, and the final handling result is output. When the evaluation indicates that the fault recovery is successful, the first processing result is output as the final handling result. The feedback reward value is calculated based on the fault recovery result corresponding to the final disposal result and the resource cost corresponding to the target disposal level during the disposal process. Using fault semantic feature vectors, target handling levels, and feedback reward values ​​as experience samples, the parameters of the feedforward neural network are updated online based on these experience samples to continuously optimize the selection of target handling levels and achieve autonomous operation and maintenance of the satellite network. The expected benefit is used to characterize the comprehensive evaluation result of the expected handling effect and resource cost when handling fault tasks at the corresponding level; the resource cost corresponding to the target handling level is used to characterize the resource overhead and latency in the handling process of the target handling level.

[0008] Secondly, the present invention provides a satellite network autonomous operation and maintenance architecture oriented towards end-edge-cloud collaboration, applicable to the satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration as described in the first aspect, wherein the satellite network autonomous operation and maintenance architecture includes: The fault semantic encoding module is used to receive natural language fault descriptions from the satellite network and extract normalized fault semantic feature vectors from the natural language fault descriptions through a pre-trained text encoding model. The intelligent hierarchical decision-making module is used to estimate the expected benefits of selecting each level in the satellite network for fault handling based on the fault semantic feature vector using a feedforward neural network, and generate the expected benefit value corresponding to each level; based on the expected benefit value corresponding to each level, through... The strategy is to select the target handling level for the fault task corresponding to the current fault information; The hierarchical handling execution module is used to send fault tasks to the corresponding nodes of the target handling level, call the fault handling engine of the target handling level to perform fault handling, and output the corresponding first processing result; The upgrade control module is used to evaluate the first processing result: when the evaluation indicates that the fault recovery has failed, a hierarchical upgrade mechanism is triggered to select the upgraded handling level; when the evaluation indicates that the fault recovery has succeeded, the first processing result is output as the final handling result. The tiered handling execution module is also used to handle fault tasks through upgraded handling levels when the assessment indicates that fault recovery has failed, until the fault recovery is successful or the highest handling level is reached, and output the final handling result. The online learning module is used to calculate feedback reward values ​​based on the fault recovery results corresponding to the final handling result and the resource costs corresponding to the target handling level during the handling process. It uses the fault semantic feature vector, the target handling level, and the feedback reward values ​​as experience samples to update the parameters of the feedforward neural network in the intelligent hierarchical decision-making module online, continuously optimizing the selection of the target handling level and achieving autonomous operation and maintenance of the satellite network. The expected benefit is used to characterize the comprehensive evaluation result of the expected handling effect and resource costs when handling fault tasks at the corresponding level; the resource costs corresponding to the target handling level are used to characterize the resource overhead and latency during the handling process at the target handling level.

[0009] The beneficial effects of this invention are: The solution provided by this invention constructs a three-layer collaborative operation system of end-edge-cloud, deploying fault handling capabilities layered to the end, edge, and cloud nodes. This effectively decentralizes fault handling capabilities to the network edge, reducing reliance on satellite-to-ground communication and fault response latency, and further improving the real-time autonomous operation and maintenance capabilities of the satellite network. Furthermore, by designing a system based on fault semantic feature vectors and... The intelligent hierarchical decision-making mechanism of the strategy can autonomously select the appropriate handling level for fault tasks based on the semantic feature vector of the fault, which is conducive to improving the accuracy of level selection and resource utilization efficiency, and further improving the system's adaptability to different types of faults. Furthermore, by establishing a unified cross-level collaboration and hierarchical upgrade mechanism, this invention encapsulates fault descriptions, executed handling action sequences, and failure / anomaly logs into a standardized operation and maintenance context. This enables continuous flow and autonomous collaborative handling of fault tasks between the edge layer, cloud layer, and terminal layer, which is conducive to improving the autonomous recovery efficiency of complex faults and further improving network operational reliability. By introducing an online learning and strategy evolution mechanism, calculating feedback reward values ​​and storing experience samples, and triggering the autonomous parameter update of a lightweight feedforward neural network, the system can continuously optimize the hierarchical decision-making strategy based on historical operation and maintenance results, which is conducive to improving the system's adaptability to unknown faults and further improving long-term autonomous operation and maintenance performance. Attached Figure Description

[0010] Figure 1 This is a schematic diagram illustrating the steps of a satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration provided in an embodiment of the present invention. Figure 3 This is a principle block diagram of a satellite network autonomous operation and maintenance architecture oriented towards end-edge-cloud collaboration provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the architecture of the edge-cloud three-layer satellite network provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the implementation of autonomous operation and maintenance of a satellite network oriented towards end-edge-cloud collaboration, as provided in an embodiment of the present invention. Figure 6 This is a comparison chart of fault handling delays between a satellite network autonomous operation and maintenance method and a ground station centralized control method provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0012] To address the problems in existing satellite network operation and maintenance processes, such as reliance on manual fault handling, insufficient collaborative capabilities across different computing levels, low efficiency in autonomous handling of complex faults, and difficulty in continuously optimizing operation and maintenance strategies, this invention provides a satellite network autonomous operation and maintenance architecture and method oriented towards end-edge-cloud collaboration.

[0013] Below, we will first introduce a satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration provided by an embodiment of the present invention.

[0014] This invention provides a method for autonomous operation and maintenance of satellite networks oriented towards end-edge-cloud collaboration, such as... Figure 1 and Figure 2 As shown, it may include the following steps: S1 acquires fault information generated during the operation of the satellite network, processes the fault information into text, and generates a natural language fault description.

[0015] Fault information may include the time of fault occurrence, fault type, a description of the fault phenomenon in natural language, and the satellite identifier of the fault.

[0016] In satellite networks, the end layer, edge layer, and cloud layer are logical functional layers: the end layer can include low-Earth orbit satellite nodes, the edge layer can include medium-Earth orbit satellite nodes, and the cloud layer can include high-Earth orbit satellites and / or ground control stations. The edge layer, cloud layer, and end layer each have different computing resources and fault handling capabilities, which together constitute the edge-cloud collaborative autonomous operation and maintenance system.

[0017] Each low-Earth orbit satellite node in the satellite network collects its own telemetry parameters and key health indicators in real time, comparing them with preset threshold values. When any telemetry parameter remains outside the normal range for an extended period exceeding the time window corresponding to the preset threshold, a fault event is triggered. Upon triggering, the abnormal telemetry parameters are combined and converted into a structured fault description. After conversion, the structured fault description, along with the fault occurrence timestamp and satellite identifier, is encapsulated into a standard fault information format to generate a natural language fault description. Understandably, the preset threshold values ​​are configurable parameters, and their specific values ​​can be flexibly determined based on the satellite platform, payload type, telemetry sampling period, and mission phase.

[0018] S2 extracts normalized fault semantic feature vectors from natural language fault descriptions using a pre-trained text encoding model.

[0019] The generated natural language fault description is input into a pre-trained text encoding model to extract fault semantic features, generating a normalized fault semantic feature vector, which provides a unified semantic representation for subsequent autonomous hierarchical decision-making. The text encoding model can be a CoSENT matching model trained on a Chinese STS training set to extract fault semantic features.

[0020] S3 processes the fault semantic feature vector through a feedforward neural network, estimates the expected benefits of fault handling at each level in the satellite network, and generates the expected benefit value corresponding to each level; the level can include the end layer, the edge layer, and the cloud layer.

[0021] This invention employs a lightweight feedforward neural network, corresponding to the expected return value of the lightweight feedforward neural network output. The expression is as follows: ; in, This represents the expected benefit value when the fault task is handled at the endpoint. This represents the expected reward value when a faulty task is handled by a side layer. This indicates the expected benefit value when the faulty task is handled by selecting the cloud layer. This represents the mapping function for a lightweight feedforward neural network. This represents the normalized semantic feature vector of the fault. Expected benefit is used to characterize the comprehensive evaluation result of the expected handling effect and resource cost when handling fault tasks at the corresponding level.

[0022] Autonomous operation and maintenance methods for satellite networks may also include: After generating the expected revenue values ​​for each level, the expected revenue values ​​are adjusted according to the current network operating status of the nodes; the network operating status includes level load and link latency.

[0023] S4, based on the expected return value corresponding to each level, through The strategy selects the target handling level for the fault task corresponding to the current fault information, and may include: With probability Select the level with the highest expected return from each level as the target disposal level, or use probability... One level is randomly selected from each level as the target treatment level, thus balancing the use of historical experience with the exploration of unknown scenarios; among them... Indicates the first The exploration rate corresponding to each training iteration, exploration rate The exploration rate gradually decreases with each training round, and is updated as follows: ; in, Indicates the first The exploration rate corresponding to each training iteration. This represents the minimum allowable exploration rate. Indicates the initial exploration rate. This represents the exploration rate decay factor. Indicates the number of training steps. This indicates taking the larger of the two values.

[0024] For example, The value can be 0.4. When selecting the target disposal level, there is a 0.6 probability to select the level with the highest expected return value from the expected return values ​​corresponding to each level as the target disposal level, and a 0.4 probability to randomly select one level from each level as the target disposal level.

[0025] S5 sends the fault task to the corresponding node at the target handling level, calls the fault handling engine at that target handling level to perform fault handling, and outputs the corresponding first processing result, which may include: When the target handling level is the end layer, the fault task is sent to the corresponding node of the end layer. The low-orbit satellite node calls the lightweight rule base, threshold strategy and fault case knowledge base to perform fast matching and handling and output the first processing result. When the target handling level is the edge layer, the fault task is sent to the corresponding node of the edge layer, and the medium-orbit satellite node calls the lightweight inference model to perform fault analysis and collaborative handling and output the first processing result; When the target handling level is the cloud layer, the fault task is sent to the corresponding node in the cloud layer, and the high-orbit satellite and / or ground center station call the large language model and fault case knowledge base to perform the final handling and output the first processing result.

[0026] S6. Evaluate the first processing result according to the preset recovery judgment rules. When the evaluation indicates that the fault recovery has failed, trigger the hierarchical escalation mechanism to select the upgraded handling level for fault task handling until the fault recovery is successful or the highest handling level is reached, and output the final handling result. When the evaluation indicates that the fault recovery is successful, output the first processing result as the final handling result.

[0027] When evaluating the first processing result, the preset recovery judgment rules may include: Whether the fault condition has been eliminated, whether system functions have returned to normal, and whether operational indicators have returned to normal. If the fault condition has not been eliminated, system functions have not returned to normal, or operational indicators have not returned to normal, the fault recovery is assessed as a failure.

[0028] When the fault recovery is deemed unsuccessful, a hierarchical escalation mechanism is triggered to select the upgraded handling level for fault task handling until fault recovery is successful or the highest handling level is reached. The final handling result is then output, which may include: When the assessment indicates that fault recovery has failed, the target handling level is unable to complete the fault handling, triggering a hierarchical escalation mechanism to raise the handling level in the order of end layer, edge layer, and cloud layer. The sequence of actions executed at the target handling level and the failure and exception logs are encapsulated into an operation and maintenance context; the operation and maintenance context is sent to the upgraded handling level for fault handling until the fault is successfully recovered or the highest handling level is reached, and the final handling result is output.

[0029] Understandably, when the target handling level is unable to handle the fault, a hierarchical escalation mechanism is automatically triggered. The fault description, the sequence of handling actions already executed at this level, and the failure exception logs are encapsulated into an operation and maintenance context and reported to a higher level for further handling until the fault is successfully recovered.

[0030] Specifically, a hierarchical upgrade mechanism may include: When the end layer is unable to complete the fault handling task, the fault description, the sequence of handling actions already executed by the end layer and the failure exception log are encapsulated into an operation and maintenance context, and reported to the satellite in the edge layer through the inter-satellite link for further handling. When the edge layer is unable to complete the fault handling task, the fault description, the sequence of handling actions already executed by the edge layer, and the failure exception log are encapsulated into an operation and maintenance context, and reported to the cloud layer for further handling via inter-satellite link or satellite-to-ground link. The operation and maintenance context is continuously transmitted during the hierarchical upgrade process.

[0031] S7. Calculate the feedback reward value based on the fault recovery result corresponding to the final disposal result and the resource cost corresponding to the target disposal level during the disposal process.

[0032] In one optional implementation, the feedback reward value is determined comprehensively based on the fault recovery result, the recovery quality evaluation value, and the resource cost corresponding to the target disposal level, and its expression is: ; in, Indicates the first The feedback reward value corresponding to the handling of this fault; Indicates the first The base reward value corresponding to the fault recovery result of the fault handling is determined by taking a positive reward value when the fault recovery is successful and a negative reward value when the fault recovery fails. Indicates the first The quality evaluation value corresponding to the fault recovery of the secondary fault handling is used to characterize the degree of system function recovery or the degree of recovery of operating indicators; Indicates the first The resource cost corresponding to the target handling level of the sub-fault handling is used to characterize the computational resource overhead, communication overhead, and response latency in the handling process of that level; , , These are weighting coefficients used to adjust the impact of fault recovery results, recovery quality, and resource costs on the feedback reward value; typically, fault recovery results are the primary evaluation factor, therefore... It can be set to greater than and The weight value can be appropriately increased when the task emphasizes low latency or low resource overhead. The value of .

[0033] The resource cost corresponding to the target disposal level can be used to characterize the resource overhead and latency in the disposal process at that level. The calculated feedback reward value can be used to evaluate the merits of the current level selection strategy and serve as the basis for subsequent online learning to update the decision network parameters.

[0034] S8 uses fault semantic feature vectors, target handling levels, and feedback reward values ​​as experience samples. Based on these experience samples, the parameters of the feedforward neural network are updated online to continuously optimize the selection of target handling levels, achieving autonomous operation and maintenance of the satellite network. This can include: The fault semantic feature vector, target handling level and feedback reward value generated during the current fault handling process are stored as an experience sample in the experience buffer. When the number of experience samples stored in the experience buffer reaches a preset threshold, the historical experience samples in the experience buffer are sampled in batches. Based on historical experience samples obtained through sampling, the parameters of the lightweight feedforward neural network are updated with the optimization objective of minimizing the error between the expected return value and the feedback reward value corresponding to the selected target treatment level, so as to continuously optimize the selection strategy of subsequent treatment levels. The preset threshold can be flexibly selected by the user between 5 and 20 according to specific needs.

[0035] Specifically, this embodiment of the invention employs an empirical replay mechanism to update and optimize the parameters of the lightweight feedforward neural network. The specific optimization process is as follows: S81, Initialize the parameters of the lightweight feedforward neural network. Initialize the experience buffer .

[0036] S82, normalizes the fault semantic feature vector generated during each fault handling process. Selected target handling level and feedback reward value As an empirical sample Store in experience buffer ; S83, Determine the experience buffer If the number of empirical samples has reached a preset threshold, then the data will be removed from the empirical buffer. A batch of empirical samples is randomly sampled, and network optimization is performed to optimize the parameters of the lightweight feedforward neural network; otherwise, return to step S82 to continue accumulating empirical samples. The network optimization process can be summarized as follows: Calculate the loss function based on the historical experience samples obtained from sampling. : ; in, Indicates batch size. Indicates the first i Fault semantic feature vectors corresponding to historical experience samples Indicates the first i The target treatment level selected from a historical experience sample This indicates that the lightweight feedforward neural network has the following parameters. Next target The output, and the selected target processing level The corresponding estimated expected return, Indicates the target handling level The corresponding feedback reward value. The parameters of the lightweight feedforward neural network are updated using gradient descent. This allows the output of the lightweight feedforward neural network to gradually approach the expected reward value.

[0037] After updating the parameters of the S84 lightweight feedforward neural network, the experience buffer is cleared, and the next round of online learning begins. Once the current fault handling is complete, subsequent fault tasks are continuously monitored and processed, enabling the system to continuously optimize hierarchical decision-making strategies based on historical fault handling experience.

[0038] Secondly, corresponding to the above method embodiments, this invention also provides a satellite network autonomous operation and maintenance architecture oriented towards end-edge-cloud collaboration, applicable to the satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration as described in the first aspect, such as... Figure 3 As shown, it may include: The system includes a fault semantic coding module, an intelligent hierarchical decision-making module, a graded handling and execution module, an escalation control module, and an online learning module. The output of the fault semantic coding module is communicatively connected to the input of the intelligent hierarchical decision module. The fault semantic coding module is used to receive natural language fault descriptions from the satellite network, extract normalized fault semantic feature vectors from the natural language fault descriptions through a pre-trained text coding model, and output the vectors to the intelligent hierarchical decision module through interface S1. Interface S1 is the information interaction interface between the fault semantic coding module and the intelligent hierarchical decision module. The input of the intelligent hierarchical decision-making module is communicatively connected to the output of the fault semantic encoding module, and the output of the intelligent hierarchical decision-making module is communicatively connected to the input of the hierarchical handling execution module. The intelligent hierarchical decision-making module is used to estimate the expected benefits of selecting each level in the satellite network for fault handling based on the fault semantic feature vector using a lightweight feedforward neural network, generating expected benefit values ​​corresponding to each level; based on the expected benefit values ​​corresponding to each level, through... The strategy selects a target handling level for the fault task corresponding to the current fault information; the selected target handling level is output to the hierarchical handling execution module through interface S2; interface S2 is the information interaction interface between the intelligent hierarchical decision module and the hierarchical handling execution module; the expected benefit is used to characterize the comprehensive evaluation result of the expected handling effect and resource cost when handling the fault task at the corresponding level; The input of the hierarchical handling execution module is communicatively connected to the output of the intelligent hierarchical decision-making module. The output of the hierarchical handling execution module is communicatively connected to the input of the upgrade control module and the input of the online learning module. The hierarchical handling execution module is used to send fault tasks to the corresponding nodes of the target handling level, call the fault handling engine of the target handling level to perform fault handling, and output the corresponding first processing result. Considering the significant differences in computing power and storage capacity among the terminal, edge, and cloud nodes, each level is configured with a handling engine of different complexity. After the handling is completed, the result is output to the upgrade control module and the online learning module through interface S3. Interface S3 is the information interaction interface between the hierarchical handling execution module and the upgrade control module and the online learning module, respectively. The upgrade control module is communicatively connected to the feedback end of the hierarchical disposal execution module. The upgrade control module is used to evaluate the first processing result: when the evaluation indicates that the fault recovery has failed, a hierarchical upgrade mechanism is triggered to select the upgraded disposal level; when the evaluation indicates that the fault recovery has succeeded, the first processing result is output as the final disposal result; the upgraded disposal level is output to the hierarchical disposal execution module through interface S4; interface S4 is the information interaction interface between the upgrade control module and the hierarchical disposal execution module.

[0039] The tiered handling execution module is also used to handle fault tasks through upgraded handling levels when the assessment indicates that fault recovery has failed, until the fault recovery is successful or the highest handling level is reached, and then output the final handling result. The input of the online learning module is communicatively connected to the output of the upgrade control module, and the output of the online learning module is communicatively connected to the feedback of the intelligent hierarchical decision-making module. The online learning module calculates the feedback reward value based on the fault recovery result corresponding to the final handling result and the resource cost corresponding to the target handling level during the handling process. It uses the fault semantic feature vector, the target handling level, and the feedback reward value as experience samples to update the parameters of the lightweight feedforward neural network in the intelligent hierarchical decision-making module online, continuously optimizing the selection of the target handling level and achieving autonomous operation and maintenance of the satellite network. The resource cost corresponding to the target handling level can be used to characterize the resource overhead and latency during the handling process of the target handling level. The online learning module outputs the updated parameters to the intelligent hierarchical decision-making module through interface S5; interface S5 is the information interaction interface between the online learning module and the intelligent hierarchical decision-making module.

[0040] This invention addresses the problems of reliance on manual labor, insufficient hierarchical collaboration, and difficulty in handling unknown faults in the operation and maintenance of large-scale satellite networks. It constructs a three-layer collaborative operation and maintenance system (end-edge-cloud) and introduces fault semantic encoding, intelligent hierarchical decision-making, graded collaborative handling, step-by-step escalation, and online learning mechanisms. The fault semantic feature vector is input into the intelligent hierarchical decision-making module, which outputs the expected benefit values ​​for the end layer, edge layer, and cloud layer through a lightweight feedforward neural network. In an optional implementation, the expected benefit value is adjusted according to the current network operating status of the node, and the target handling level is selected based on the expected benefit value. Simultaneously, online learning enables continuous optimization of the hierarchical decision-making strategy, effectively reducing reliance on manual operation and maintenance and improving the efficiency of satellite network fault handling and autonomous operation and maintenance capabilities.

[0041] A schematic diagram of the architecture of a three-layer satellite network (end-edge-cloud), as shown below. Figure 4As shown, the three-layer satellite network architecture of the present invention, consisting of end-edge-cloud, is divided into three logical layers from bottom to top. The end layer comprises numerous low-Earth orbit (LEO) satellites with high operating speed but limited onboard computing resources and storage capacity, suitable for handling faults with small impact areas and simple structures. The edge layer comprises relatively fewer medium-Earth orbit (MEO) satellites with superior computing and storage capabilities compared to end-layer nodes, capable of deploying lightweight language models, suitable for fault diagnosis and handling of moderate complexity. The cloud layer comprises high-Earth orbit (GEO) satellites and / or ground control stations. Compared to the end and edge layers, the cloud layer possesses stronger computing, storage, global knowledge aggregation, and complex reasoning capabilities, capable of deploying complex algorithms such as large language models, used to handle complex faults with large impact areas requiring global information. The three layers communicate with each other via inter-satellite links and satellite-to-ground links, forming a collaborative communication system between the end, edge, and cloud layers. Fault tasks can flow between different layers based on hierarchical decision results and generate feedback information after fault recovery, providing data support for subsequent online learning.

[0042] In an optional embodiment of the present invention, the hierarchical handling execution module may include an end-layer handling engine unit, an edge-layer handling engine unit, and a cloud-layer handling engine unit. The end-layer handling engine unit is deployed on a low-Earth orbit satellite and uses a rule-based and threshold-based engine to achieve rapid fault identification. The edge-layer handling engine unit is deployed on a medium-Earth orbit satellite and uses a finely tuned fault diagnosis model or a medium-scale language reasoning model to diagnose and handle medium-complexity faults. The cloud-layer handling engine unit is deployed on a high-Earth orbit satellite and / or a ground control station. Compared to the end-layer and edge-layer, the cloud layer possesses stronger computing power, storage capacity, global knowledge aggregation capability, and complex reasoning capability. Therefore, in this embodiment of the present invention, the cloud layer invokes a large-scale language model and combines it with a fault case knowledge base to perform global analysis and final handling decisions for highly complex faults.

[0043] The upgrade control module encapsulates the fault description, the sequence of actions performed at the current level, and the failure and exception logs into an operation and maintenance context, and reports this operation and maintenance context to a higher level for further processing via inter-satellite links or satellite-to-ground links.

[0044] In an optional embodiment of the present invention, the online learning module may include an experience replay buffer unit and a parameter update unit; wherein, the experience replay buffer unit is used to store experience samples consisting of fault semantic feature vectors, selected target handling levels, and feedback reward values, and the parameter update unit is used to perform batch sampling of historical experience samples when the number of samples in the experience replay buffer unit reaches a preset threshold, with the optimization objective of minimizing the error between the expected return value and the feedback reward value corresponding to the selected target handling level, and update the parameters of the lightweight feedforward neural network.

[0045] The flowchart for the implementation of autonomous operation and maintenance of satellite networks provided in this embodiment of the invention is as follows: Figure 5 As shown, a specific application scenario of the above-mentioned autonomous operation and maintenance method for satellite networks is further illustrated. The fault semantic feature vector obtained after processing by the fault semantic encoding module is first input into the intelligent hierarchical decision-making module, which outputs the expected benefit values ​​corresponding to the end layer, edge layer, and cloud layer through a lightweight feedforward neural network. In an optional implementation, the expected benefit values ​​corresponding to each layer are adjusted according to the current network operation status of the node, and the target handling layer is selected based on the expected benefit value. The fault task is then sent to the target handling layer for handling.

[0046] If the initial selection is the end layer, the low-Earth orbit satellite from the source of the fault will perform rapid handling based on a rule-based and threshold-based engine. If the fault is successfully recovered, the process is closed; otherwise, a step-by-step escalation is performed. The fault description, end-layer handling records, and failure reasons are encapsulated into an operational context and reported to the mid-Earth orbit satellite in the edge layer via inter-satellite links for further processing.

[0047] The operational context reported by the edge layer receiver is used by the medium-Earth orbit satellite to execute incremental processing using a medium-sized language inference model. If the processing is successful, the process is closed; if the edge layer still cannot complete fault recovery, the system continues to migrate the complete operational context to the cloud layer to achieve cross-layer autonomous collaborative processing.

[0048] The cloud layer receives the complete operation and maintenance context, and the high-orbit satellite and / or ground center station invoke the large language model and fault case knowledge base to perform the final handling. As the highest level, the cloud layer is responsible for handling complex faults that the end layer and edge layer cannot resolve. After the cloud layer completes the fault recovery, it feeds back the complete handling results to the autonomous operation and maintenance management module, entering the feedback optimization phase.

[0049] After a fault task is completed, the handling results, execution logs, hierarchical decision records, and evaluation indicators generated at each level are uniformly transmitted back to form a complete operational feedback. This embodiment of the invention can store the mapping information or handling experience of this fault in a knowledge base for rapid matching and reference in subsequent similar faults; then, the handling experience is input into the online learning module to update and optimize the parameters of the lightweight feedforward neural network in the intelligent hierarchical decision module, enabling the system to continuously optimize the initial hierarchical selection strategy for future similar faults based on historical feedback.

[0050] This invention, through continuous feedback learning and model optimization, gradually forms an autonomous operation and maintenance closed loop covering the entire process of fault semantic understanding, autonomous hierarchical decision-making, cross-level collaborative handling, autonomous step-by-step escalation, and online feedback optimization. This enables autonomous collaborative management and control of fault tasks across the edge, cloud, and terminal layers, and continuously improves the autonomous operation and maintenance capabilities of satellite networks. This invention uses autonomous fault handling as a typical application scenario for illustration; however, this satellite network autonomous operation and maintenance method is also applicable to other autonomous operation and maintenance tasks such as satellite network configuration anomalies, resource scheduling anomalies, and link anomalies.

[0051] To verify the effectiveness of the satellite network autonomous operation and maintenance architecture and method for end-edge-cloud collaboration provided in this embodiment of the invention, the following simulation experiments were conducted.

[0052] Simulation experimental conditions: The simulation experiment platform of this invention embodiment is: Windows 11 operating system, Python 3.9, STK (Systems Tool Kit) 12.

[0053] The simulation scenario constructs a three-layer satellite network: edge, cloud, and terminal. The terminal layer consists of 64 low-Earth orbit (LEO) satellites, the edge layer consists of 16 medium-Earth orbit (MEO) satellites, and the cloud layer includes 3 geostationary orbit (GEO) satellites and 5 ground control stations. Fault events are generated randomly, with arrival times following a homogeneous Poisson distribution. The average fault occurrence rate is set to 0.219 events per hour. Fault types include power supply and distribution faults, propulsion faults, attitude control faults, telemetry and control (TT&C) faults, thermal control faults, and structural and mechanical faults. Fault descriptions are represented in natural language, and a fault case knowledge base is constructed based on corresponding standard handling strategies to simulate the autonomous operation and maintenance process under different fault scenarios. The simulation constellation parameters are shown in Table 1.

[0054] Table 1 Simulation Constellation Parameters

[0055] Simulation content: Simulation 1: A comparison of fault handling latency between the edge-cloud collaborative autonomous operation and maintenance method provided in this embodiment of the invention and the traditional ground station centralized control method, as shown in the figure. Figure 6 As shown.

[0056] Depend on Figure 6 It is evident that the satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration proposed in this invention has a lower fault response latency compared to the traditional centralized control method of ground stations. Traditional methods are limited by the satellite-to-ground visibility window, resulting in a relatively dispersed distribution of fault handling latency. In the constructed satellite network fault operation and maintenance simulation scenario, the average handling latency, including simulation experimental data, is approximately 5335 seconds (about 1.5 hours). This invention, by deploying fault diagnosis and handling capabilities at the end-layer and edge-layer nodes, enables most simple or moderately complex faults to be handled at the end-layer or edge-layer, with an average fault handling latency of approximately 12 seconds. Only in the case of complex faults is the fault escalated to the cloud layer. Therefore, this invention effectively reduces fault response time, verifying the effectiveness of the proposed end-edge-cloud collaborative satellite network autonomous operation and maintenance architecture.

[0057] Simulation 2: In the simulation experiment of this embodiment, the full edge layer processing method, the full cloud layer processing method and the end-edge-cloud collaborative autonomous operation and maintenance method proposed in this invention are simulated respectively. The failure handling success rate and average end-to-end latency of the three methods are compared. The results are shown in Table 2.

[0058] Table 2 Performance Comparison of Different Fault Handling Methods

[0059] As shown in Table 2, the satellite network autonomous operation and maintenance method proposed in this embodiment achieves a better balance between fault handling success rate and end-to-end latency compared to the full-edge-layer processing method and the full-cloud-layer processing method. While the full-edge-layer processing method has lower latency, its success rate for handling complex faults is lower due to the limited capabilities of the edge-layer model. Although the full-cloud-layer processing method can utilize a large language model to perform complex fault reasoning, the cross-layer transmission and satellite-to-ground communication processes increase the overall response latency. This embodiment of the invention, through an intelligent hierarchical decision-making mechanism, autonomously selects the initial handling level based on the semantic features of the fault, and combines this with a step-by-step escalation mechanism. The step-by-step escalation mechanism is only triggered and higher-level resources are invoked when fault recovery cannot be completed at the current level. Therefore, it balances fault recovery success rate and real-time performance, verifying the effectiveness of the hierarchical collaborative handling mechanism proposed in this embodiment.

[0060] In summary, the autonomous operation and maintenance architecture and method for satellite networks oriented towards end-edge-cloud collaboration provided by this invention has the following beneficial effects: This invention proposes an autonomous operation and maintenance architecture for satellite networks that is oriented towards end-edge-cloud collaboration. It unifies functions such as fault semantic understanding, intelligent hierarchical decision-making, graded handling and execution, cross-level upgrades, and online learning, and realizes the collaborative deployment of autonomous operation and maintenance capabilities among the end layer, edge layer, and cloud layer. This forms a complete closed loop of autonomous fault perception, autonomous decision-making, autonomous handling, and autonomous optimization, thereby improving the autonomous operation capability of satellite networks.

[0061] This invention embodiment designs based on fault semantic features and The intelligent hierarchical decision-making mechanism of the strategy can autonomously select the appropriate handling level for the fault task based on the semantic characteristics of the fault, which is conducive to improving the accuracy of level selection and resource utilization efficiency, and further improving the system's adaptability to different types of faults.

[0062] This invention establishes a unified cross-level collaboration and step-by-step upgrade mechanism, encapsulating the original fault description, the sequence of executed handling actions, and the failure and anomaly logs into a standardized operation and maintenance context. This enables the continuous flow and autonomous collaborative handling of fault tasks between the end layer, edge layer, and cloud layer, which is beneficial to improving the autonomous recovery efficiency of complex faults and further enhancing the reliability of network operation.

[0063] By introducing an online learning and strategy evolution mechanism, this invention continuously updates the parameters of a lightweight feedforward neural network based on an experience buffer. This enables the system to continuously optimize hierarchical decision-making strategies based on historical operation and maintenance results, which helps improve the system's adaptability to unknown faults and further enhances its long-term autonomous operation and maintenance performance.

[0064] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for autonomous operation and maintenance of satellite networks oriented towards end-edge-cloud collaboration, characterized in that, include: The system acquires fault information generated during the operation of the satellite network, performs textual processing on the fault information, and generates natural language fault descriptions. Normalized fault semantic feature vectors are extracted from the natural language fault description using a pre-trained text encoding model. The fault semantic feature vector is processed by a feedforward neural network to estimate the expected benefits of fault handling at each level in the satellite network and generate the expected benefit value corresponding to each level; the levels include the end layer, the edge layer and the cloud layer. Based on the expected returns corresponding to each level, through The strategy is to select the target handling level for the fault task corresponding to the current fault information; The fault task is sent to the corresponding node of the target handling level, the fault handling engine of the target handling level is invoked to perform fault handling, and the corresponding first processing result is output. The first processing result is evaluated according to the preset recovery judgment rules. When the evaluation indicates that the fault recovery has failed, a hierarchical escalation mechanism is triggered to select the upgraded handling level for fault task handling until the fault recovery is successful or the highest handling level is reached, and the final handling result is output. When the fault recovery is deemed successful, the first processing result will be output as the final disposal result. The feedback reward value is calculated based on the fault recovery result corresponding to the final disposal result and the resource cost corresponding to the target disposal level during the disposal process. Using fault semantic feature vectors, target handling levels, and feedback reward values ​​as experience samples, the parameters of the feedforward neural network are updated online based on these experience samples to continuously optimize the selection of target handling levels and achieve autonomous operation and maintenance of the satellite network. The expected benefit is used to characterize the comprehensive evaluation result of the expected handling effect and resource cost when handling fault tasks at the corresponding level; the resource cost corresponding to the target handling level is used to characterize the resource overhead and latency in the handling process of the target handling level.

2. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, The end layer includes low-Earth orbit satellite nodes, the side layer includes medium-Earth orbit satellite nodes, and the cloud layer includes high-Earth orbit satellites and / or ground control stations. The endpoint, edge, and cloud layers each possess different computing resources and fault handling capabilities, collectively forming an edge-cloud collaborative autonomous operation and maintenance system.

3. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, Autonomous operation and maintenance methods for satellite networks also include: After generating the expected revenue values ​​for each level, the expected revenue values ​​are adjusted according to the current network operating status of the node; the network operating status includes level load and link latency.

4. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, Based on the expected returns corresponding to each level, through The strategy selects the target handling level for the fault task corresponding to the current fault information, including: With probability Select the level with the highest expected return from each level as the target disposal level, or use probability... One level is randomly selected from each level as the target treatment level, thus balancing the use of historical experience with the exploration of unknown scenarios; among them... Indicates the first The exploration rate corresponding to each training iteration, exploration rate The exploration rate gradually decreases with each training round, and is updated as follows: ； in, Indicates the first The exploration rate corresponding to each training iteration. This represents the minimum allowable exploration rate. Indicates the initial exploration rate. This represents the exploration rate decay factor. Indicates the number of training steps. This indicates taking the larger of the two values.

5. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 2, characterized in that, The process of sending a fault task to the corresponding node at the target handling level, invoking the fault handling engine at that target handling level to perform fault handling, and outputting the corresponding first processing result includes: When the target handling level is the end layer, the fault task is sent to the corresponding node of the end layer. The low-orbit satellite node calls the rule base, threshold strategy and fault case knowledge base to perform fast matching and handling and output the first processing result. When the target handling level is the edge layer, the fault task is sent to the corresponding node of the edge layer, and the medium-orbit satellite node calls the inference model to perform fault analysis and collaborative handling and output the first processing result; When the target handling level is the cloud layer, the fault task is sent to the corresponding node in the cloud layer, and the high-orbit satellite and / or ground center station call the large language model and fault case knowledge base to perform the final handling and output the first processing result.

6. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, When evaluating the first processing result, the preset recovery judgment rules include: Whether the fault condition has been eliminated, whether the system functions have returned to normal, and whether the operating indicators have returned to normal.

7. The satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, When the fault recovery is deemed unsuccessful, a hierarchical escalation mechanism is triggered to select the next higher level of handling for the fault task, continuing until fault recovery is successful or the highest handling level is reached. The final handling result is then output, including: When the assessment indicates that fault recovery has failed, the target handling level is unable to complete the fault handling, triggering a hierarchical escalation mechanism to raise the handling level in the order of end layer, edge layer, and cloud layer. The sequence of actions executed at the target handling level and the failure and exception logs are encapsulated into an operation and maintenance context; the operation and maintenance context is sent to the upgraded handling level for fault handling until the fault is successfully recovered or the highest handling level is reached, and the corresponding final handling result is output.

8. The autonomous operation and maintenance method for satellite networks oriented towards end-edge-cloud collaboration according to claim 1, characterized in that, The process of continuously optimizing the selection of target treatment levels by updating the parameters of the feedforward neural network online based on the empirical samples includes: The fault semantic feature vector, target handling level and feedback reward value generated during the current fault handling process are stored as an experience sample in the experience buffer. When the number of experience samples stored in the experience buffer reaches a preset threshold, the historical experience samples in the experience buffer are sampled in batches. Based on the historical experience samples obtained from sampling, the parameters of the feedforward neural network are updated with the optimization objective of minimizing the error between the expected return value and the feedback reward value corresponding to the selected target treatment level, so as to continuously optimize the selection strategy of subsequent treatment levels.

9. A satellite network autonomous operation and maintenance architecture oriented towards end-edge-cloud collaboration, applicable to the satellite network autonomous operation and maintenance method oriented towards end-edge-cloud collaboration as described in any one of claims 1 to 8, characterized in that, The autonomous operation and maintenance architecture of the satellite network includes: The fault semantic encoding module is used to receive natural language fault descriptions from the satellite network and extract normalized fault semantic feature vectors from the natural language fault descriptions through a pre-trained text encoding model. The intelligent hierarchical decision-making module is used to estimate the expected benefits of selecting each level in the satellite network for fault handling based on the fault semantic feature vector using a feedforward neural network, and generate the expected benefit value corresponding to each level; based on the expected benefit value corresponding to each level, through... The strategy is to select the target handling level for the fault task corresponding to the current fault information; The hierarchical handling execution module is used to send fault tasks to the corresponding nodes of the target handling level, call the fault handling engine of the target handling level to perform fault handling, and output the corresponding first processing result; The upgrade control module is used to evaluate the first processing result: when the evaluation indicates that the fault recovery has failed, a hierarchical upgrade mechanism is triggered to select the upgraded handling level; when the evaluation indicates that the fault recovery has succeeded, the first processing result is output as the final handling result. The tiered handling execution module is also used to handle fault tasks through upgraded handling levels when the assessment indicates that fault recovery has failed, until the fault recovery is successful or the highest handling level is reached, and output the final handling result. The online learning module is used to calculate the feedback reward value based on the fault recovery result corresponding to the final handling result and the resource cost corresponding to the target handling level during the handling process. The fault semantic feature vector, the target handling level and the feedback reward value are used as experience samples. Based on the experience samples, the parameters of the feedforward neural network in the intelligent hierarchical decision module are updated online to continuously optimize the selection of the target handling level and realize the autonomous operation and maintenance of the satellite network. Among them, the resource cost corresponding to the target handling level is used to characterize the resource overhead and latency in the handling process of the target handling level.

Images