Satellite communication network construction method based on big language model reasoning

Abstract

The invention discloses a satellite communication network construction method based on large language model reasoning, and relates to the technical field of satellite communication and intelligent network control, and the method comprises the steps: carrying out the deep semantic analysis of a high-level task instruction through a large language model engine, converting a fuzzy natural language demand into an accurate executable strategy framework, and carrying out the deep semantic analysis of a high-level task instruction; the fundamental transformation from how to configure network management and control to what is achieved is realized, and the intelligent level and the man-machine interaction efficiency of network management are greatly improved; a space-time diagram model is introduced to carry out prospective prediction on network topology, and a comprehensive link cost value function and a multi-target routing evaluation function are designed, so that a system can accurately predict and avoid network faults caused by cross joints and the like, the problem of communication continuity in special areas such as polar regions and the like is fundamentally solved, and the communication efficiency is improved. Excessive dependence on ground relay facilities is eliminated, and end-to-end transmission delay is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite communication and intelligent network control, and in particular to a satellite communication network construction method based on large language model reasoning. BACKGROUND

[0002] Currently, low earth orbit satellite constellations have become the core means of constructing global coverage communication networks. Such networks rely on a large number of satellites working together to form a space information transmission backbone network through inter-satellite links, aiming to provide continuous services to global regions including the polar regions. The polar regions have important value in the fields of climate research, channel monitoring, and emergency communication due to their special geographical location. Achieving stable and reliable low-latency communication coverage is a basic goal of satellite network construction.

[0003] In the process of deploying a satellite network in the polar region, the satellite constellations form an intersection joint in the intersection area of the orbital planes near the polar region. In this area, there is high-speed relative motion between adjacent orbital planes. This motion characteristic makes it impossible to establish or maintain stable inter-satellite link connections between satellites, causing the network topology to appear periodic disconnection. If the data transmission path passes through this area, it will face the risk of connection interruption and data packet loss, affecting the continuity of polar communication services.

[0004] To meet the communication needs of the polar region, some existing solutions introduce ground gateway stations as relay nodes to guide polar region traffic through ground networks for forwarding. This method compensates for the defects of the space network topology through ground infrastructure to ensure that the communication connection does not interrupt. However, this method lengthens the data propagation path, as data packets need to be transmitted from the satellite downlink to the ground station, transmitted through the ground network, and then uplinked to the satellite, increasing the end-to-end transmission delay. At the same time, the implementation of such a solution relies on the construction and maintenance of ground stations in specific geographic areas. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a satellite communication network construction method based on large language model reasoning to solve the problem of periodic interruption of inter-satellite links caused by the intersection joint generated by the orbital motion of satellites in the polar region and other areas. The existing solution that relies on ground relays lengthens the delay and is costly.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides a satellite communication network construction method based on large language model reasoning, which includes,

[0009] Step S1, deploying a satellite network cognitive decision hub containing a large language model inference engine on the ground or on a satellite, which interacts with a topology prediction submodule, a path search submodule, and a strategy verification submodule via a structured tool interface;

[0010] Step S2, receiving multi-modal input of network state data and high-level task instructions, which include service level objectives and constraints;

[0011] Step S3, performing semantic analysis and intent extraction on the multi-modal input by the large language model inference engine, generating a network configuration strategy framework containing path intent, resource and strategy constraints, and target SLA;

[0012] Step S4, calling the topology prediction submodule to generate a spatio-temporal graph model for the target period and performing consistency and safety verification by the strategy verification submodule;

[0013] Step S5, mapping the strategy framework to application programming interface call instructions for a software-defined network controller through a rule converter;

[0014] Step S6, issuing network control instructions to related nodes of the satellite communication network by the software-defined network controller, updating flow table entries to build and activate the target path.

[0015] As a preferred scheme of the satellite communication network construction method based on large language model inference, the network state data includes at least one of the following: satellite orbit ephemeris data, inter-satellite link state data, load and availability data of ground gateways and border gateways, real-time or forecast weather data, and historical routing and alarm logs.

[0016] As a preferred scheme of the satellite communication network construction method based on large language model inference, the analysis of high-level task instructions by the large language model inference engine includes:

[0017] Extracting task objectives, target areas, service level objectives, and mandatory constraints and binding the extraction results as machine-readable strategy constraints.

[0018] As a preferred scheme of the satellite communication network construction method based on large language model inference, the topology prediction submodule is constructed as a spatio-temporal graph model with satellites and ground gateways as nodes, inter-satellite links and satellite-ground links as edges, and link weights varying with time, and the predicted failed links and their effective time windows are labeled; where link failure includes cross-joint area interruption caused by satellite relative motion.

[0019] The time function of the link weight is defined as follows:

[0020] The visibility, slant range and elevation angle of the link are obtained based on ephemeris and station data, and a predicted failure time window of the cross joint is generated, and the disable list and the maintenance window are integrated into the same availability timeline;

[0021] The original index sequences of delay, jitter, packet loss and available bandwidth are obtained by combining the frequency band and power link budget, weather attenuation and link load, and the representative values at the current time are extracted by sliding window;

[0022] Each original index is mapped to a cost type quantity in [0, 1] and is upper and lower limit clipped, and the mapping process is:

[0023] When the direction of a certain index is inconsistent, that is, the greater the worse or the greater the better, the direction factor is processed first, and then linear segmented clipping is performed, so that indexes with different dimensions can be compared;

[0024]

[0025] Wherein, y e,j (t) represents the corresponding original index, L j ,U j respectively represent the engineering lower limit and upper limit of the index, s j is a direction factor, which is 1, indicating that the greater the value, the greater the cost, and -1, indicating that the smaller the value, the greater the cost;

[0026] The outlier suppression and exponential smoothing are applied to , the time constant is set according to the business timeliness, so as to reduce the influence of short-time jitter on route switching;

[0027] Under the availability gating, the link cost is obtained by weighted aggregation, and a time proximity penalty is added to the link near the failure window. When in the failure window or disabled by strategy, the weight is set to a maximum value to prohibit selection:

[0028]

[0029] Wherein, ω e (t) represents the time variation weight of link e at time t, which is dimensionless, and the smaller the value, the better, χ e (t) is an availability indicator, which is 1 when in the predicted failure window or disabled, otherwise 0, M represents the number of indicators participating in aggregation, a j is the non-negative aggregation weight of the jth index, κ is the failure proximity penalty coefficient, is the starting time of the next failure window of link e, Δ is the advance amount, which is used to raise the cost before failure, Ω is a maximum constant, which is used to represent the cost of prohibiting routing, and max(·) represents the upper bound of the interval;

[0030] Re-calculate ω in fixed beat or event triggered manner e (t), the events include link state change, weather update and load mutation, the weight update enters a path search submodule, and is used for multi-objective evaluation function routing and preset switching.

[0031] As a preferred scheme of the satellite communication network construction method based on large language model reasoning, wherein: the path search submodule calculates a candidate path set meeting the strategy constraint for the data flow to be forwarded in the space-time graph model, and selects a target path according to a multi-objective routing evaluation function, the target path is preset before the data flow reaches the predicted failure link and is activated when the trigger condition is met;

[0032] The multi-objective routing evaluation function includes:

[0033] Establish an evaluation dimension set And the candidate path set is gated by the forced constraint of the strategy verification submodule, the path violating the disabled list or the regulatory domain is directly excluded, and does not enter the scoring;

[0034] For each candidate path p, a predicted path-level indicator is obtained at time t The sum type adopts link summation, the bottleneck type adopts minimum value or maximum value, and the statistical caliber is consistent with the network state data;

[0035] After the gating passes, the path score is calculated in the weighted form of structure cost and SLA deviation penalty:

[0036]

[0037] Wherein, J p (t; f) represents the comprehensive score of data flow f through path p at time t, the smaller the value is, the better, λ ω , λ sla is the weight of structure cost and SLA penalty, the value is 0-1, and λ ω + λ sla = 1, E p is the link set included in path p, |E p | is the cardinality thereof, unit: piece, ω e (t) is the time variation weight of link e at time t, dimensionless, is the evaluation dimension set, b m (f) is the normalized weight of data flow f in dimension m, non-negative and the sum of each item is 1, max(·) is the maximum value operator, r m is the direction factor, 1 for delay / jitter / packet loss, -1 for throughput / availability, is the path-level predicted value, corresponding target value, σ m >0 is a normalized scale of dimension m;

[0038] J p (t; f) is time-smoothed to suppress instantaneous jumps, and the remaining bandwidth is prioritized in terms of shorter path and further failure window;

[0039] A priority vector ρ m (f) is extracted from high-level task instructions and normalized as evaluation weights:

[0040]

[0041] where b m (f) is a normalized weight of dimension m, ρ m (f) is the original priority derived from service level target, non-negative;

[0042] The path with the smallest J p (t; f) is selected as the target path in the candidate set, and the suboptimal path is formed into a preset alternative according to the score gap for triggering switching.

[0043] As a preferred scheme of the satellite communication network construction method based on large language model reasoning, the activation includes: completing the pre-installation of the target path before the effective window of the predicted failed link, and triggering the condition based on the arrival time prediction, the link state change event or the threshold alarm, and when triggered, performing sequential activation on the old path and the new path to reduce packet loss and reordering.

[0044] As a preferred scheme of the satellite communication network construction method based on large language model reasoning, the rule converter maps the natural language description of the path intention and the policy constraint to the instruction set for the software-defined network controller.

[0045] The instruction set includes: flow table addition, deletion, modification, path intention arrangement, batch transaction delivery, versioning and rollback control, idempotent verification and permission audit.

[0046] As a preferred scheme of the satellite communication network construction method based on large language model reasoning, the network control instruction is used to update the flow table entry of each node along the target path.

[0047] The flow table entry includes at least a matching field and an action field,

[0048] The matching field includes one of five-tuple, service identifier or label identifier;

[0049] The action field includes one or more of forwarding port, next hop identifier, label stack operation and queue priority setting.

[0050] As a preferred scheme of the satellite communication network construction method based on large language model reasoning provided in the application, the predicted failed link comprises at least one of the following:

[0051] Adjacent orbital plane inter-satellite links of the cross joint area, satellite-ground links with reduced availability due to elevation angle or weather, and temporarily unavailable links due to link congestion or maintenance window.

[0052] As a preferred scheme of the satellite communication network construction method based on large language model reasoning provided in the application, the strategy verification submodule performs constraint consistency verification and simulation verification on the to-be-issued strategy, triggers automatic revision or rollback when the verification fails, and performs closed-loop evaluation on the service level target after the strategy takes effect and feeds back the evaluation result to the large language model reasoning engine to optimize subsequent decision-making.

[0053] The application has the following advantages: the large language model engine is used to perform deep semantic analysis on high-level task instructions, convert fuzzy natural language requirements into accurate executable strategy framework, realize a fundamental change from how to configure to what to achieve, and greatly improve the intelligent level of network management and the efficiency of human-computer interaction. Secondly, the spatiotemporal graph model is introduced to make forward-looking prediction on the network topology, and a comprehensive link value function and a multi-objective routing evaluation function are designed, so that the system can accurately predict and avoid network faults caused by cross joints, fundamentally solve the communication continuity problem in special areas such as the polar region, and reduce the dependence on ground relay facilities, effectively reducing the end-to-end transmission delay.

[0054] The preset-trigger mechanism is used to realize the smooth switching of the service path without feeling, avoiding packet loss and severe jitter in the traditional reactive routing convergence process, and significantly improving the user experience and the reliability of key services. The introduction of the rule converter and the strategy verification submodule ensures the safe, accurate and consistent mapping of high-level strategies to low-level instructions, forms a complete closed loop from intention generation to strategy execution and effect evaluation, and enhances the autonomy and anti-disturbance ability of the network. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope of the present application.

[0056] Figure 1 The flowchart of the satellite communication network construction method in the embodiments. DETAILED DESCRIPTION

[0057] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0058] All the terms used in the present application (including technical and scientific terms) have the meanings generally understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0059] For example, the terms "first", "second", etc. used in the present application are only used to distinguish the description of similar objects, to distinguish the first object from another object, and are not used to describe a specific order or sequence, nor can be understood as indicating or implying relative importance.

[0060] The present application proposes a satellite communication network construction method based on large language model reasoning, combining Figure 1 As shown in the figure, the method comprises:

[0061] Step S1, deploying a satellite network cognitive decision hub containing a large language model reasoning engine on the ground or in space, the decision hub interacting with a topology prediction submodule, a path search submodule and a strategy verification submodule via a structured tool interface;

[0062] In the present embodiment, the so-called structured tool interface refers to a fixed message mode and call convention for machines, which is used to decouple the interaction of the reasoning engine and the three submodules, and the input and output are serializable key-value sets or equivalent object representations. Further, the topology prediction submodule, the path search submodule and the strategy verification submodule each only undertake three types of functions, namely geometry and availability prediction, path candidate and evaluation, and constraint consistency and safety verification, without overlapping the functions of the others, in order to ensure the single function and replaceability. The default interface polling period is 1000 ms, which can be adjusted to 200-5000 ms, the interface timeout is 500 ms by default, which can be adjusted to 200-2000 ms, the number of retries is 2 times by default, which can be adjusted to 0-5 times, and the value is based on the stability of the end-to-end control loop and the inflection point of the message congestion rate curve in offline simulation; if a submodule is unresponsive for a short time, the system executes the default degradation path: stop launching new requests and retain the latest verified strategy version, and then synchronize the differences after the interface is restored, to ensure continuous minimum availability.

[0063] Step S2, receiving multi-modal input of network state data and high-level task instructions, the high-level task instructions including service level targets and constraints;

[0064] Specifically, the multi-modal input refers to a composite load containing both structured state data and textual task instructions, where the state data includes ephemeris, links, loads, and weather, and the instruction load includes service level objectives and mandatory constraint machine-readable fragments. The numerical caliber is as follows: the ephemeris update period is 60 s by default, and can be adjusted to 10-300 s; the inter-satellite link state refresh period is 1 s by default, and can be adjusted to 200-2000 ms; the ground gateway load refresh period is 5 s by default, and can be adjusted to 1-30 s; the weather data window is 15 min by default, and can be adjusted to 5-60 min; the above update frequencies are determined according to the natural change time scale of each source and the link budget sensitivity. If any data source is missing for more than two update periods, the latest reliable sample is maintained and marked as degraded, and the weight influence of the source is reduced in strategy verification until fresh data is restored.

[0065] Step S3, the multi-modal input is subjected to semantic analysis and intent extraction by a large language model reasoning engine to generate a network configuration policy framework containing path intent, resource and policy constraints, and target SLA;

[0066] For example, the path intent is used to describe the target connectivity relationship and emphasis of the source / destination or service identifier, the resource and policy constraints are used to limit the available link set, the upper limit of energy consumption, and the regulatory domain boundary, and the target SLA is used to bind the expected interval of delay, jitter, packet loss, availability, and throughput; the analysis output is a fixed field set, and the field key name is frozen before going online for rule converter consumption. The maximum length of analysis and rewriting is 8k characters by default, and can be adjusted to 2-16k; when the intent analysis exceeds the length, the tail truncation and key field priority reservation policy is adopted, and the key field priority is executed in the order of mandatory constraint>connectivity>SLA target, so as to avoid the unexecutable configuration caused by missing items.

[0067] Step S4, a topology prediction submodule is called to generate a spatio-temporal graph model for the target period and perform consistency and safety verification through a strategy verification submodule;

[0068] Step S5, a rule converter is used to map the policy framework into application programming interface call instructions of a software defined network controller;

[0069] Step S6, the software defined network controller issues network control instructions to related nodes of the satellite communication network to update flow table entries to build and activate the target path;

[0070] In one embodiment, the network state data includes at least one of the following: satellite orbit ephemeris data, inter-satellite link state data, load and availability data of ground gateway and border gateway, real-time or forecast weather data, historical routing and alarm logs;

[0071] In one embodiment, the analysis of the large language model reasoning engine on the high-level task instruction includes:

[0072] extracting the task target, target area, service level target and mandatory constraints and binding the extraction results as machine-readable policy constraints;

[0073] The service level target is one or more of latency, jitter, packet loss rate, availability or throughput;

[0074] The mandatory constraints include a list of disabled links, an upper limit of energy consumption, a regulatory compliance domain, etc.

[0075] In one embodiment, the topology prediction submodule is constructed as a space-time graph model with satellites and ground gateways as nodes, inter-satellite links and satellite-ground links as edges, and link weights varying with time, and the predicted failed links and their effective time windows are labeled; wherein the link failure includes the interruption in the cross-seam region caused by the relative motion of the satellite;

[0076] Similarly, the time-varying of link availability and cost relies on the joint representation of geometric, physical and traffic information, and is implemented on a unified time axis; the default recalculation period is 1s, which can be adjusted to 200ms-5s; the default sliding representative value window is 3s, which can be adjusted to 1-10s; the default advance of failure proximity penalty is 30s, which can be adjusted to 10-120s; the prediction of failure window comes from ephemeris and station geometry and is superimposed with maintenance plan and disabled list. For the direction unification, normalization and robust processing in the cost function, the acquisition method is: configure the direction factor according to the engineering meaning that the larger the index is, the better / the worse it is, and complete the interval clipping with upper and lower limits, use quantile threshold or equivalent robust scale for outlier suppression, and select the exponential smoothing time constant so that the switching jitter is within the tolerance range of the business. Optionally, if the geometry changes rapidly, the recalculation period and the smoothing time constant are temporarily increased to suppress the jitter, but the gating semantics remains unchanged. If the meteorological or load source is missing, only the geometric and historical statistical items are used to generate the cost baseline, and the value of the corresponding edge is forced to be raised until fresh samples arrive.

[0077] The time function of the link weight is defined as follows:

[0078] Based on ephemeris and station data, the visibility, slant range and elevation angle of the link are obtained, and the predicted failure time window of the cross-seam is generated, and the disabled list and maintenance window are integrated into the same availability time axis;

[0079] Combined with frequency band and power link budget, weather attenuation and link load, the original index sequence of delay, jitter, packet loss, available bandwidth, etc. is obtained, and the representative value at the current time is extracted with a sliding window;

[0080] Each original index is mapped to a cost type quantity in [0, 1] and is clipped with upper and lower limits, and the mapping process is:

[0081] When the greater the worse or the greater the better direction is inconsistent, first handle the direction factor, and then do linear segmentation clipping, so that indicators of different dimensions can be compared;

[0082]

[0083] wherein, yj (t) represents the jth normalized cost of link e at time t, taking value [0, 1], y e,j (t) represents the corresponding original index (such as delay, packet loss, utilization, etc.), L j U j respectively represent the engineering lower limit and upper limit of the index, s j is a direction factor, taking 1 to represent that the greater the value, the greater the cost, and taking -1 to represent that the smaller the value, the greater the cost;

[0084] On outlier suppression and exponential smoothing are applied, and the time constant is set according to the timeliness of the business to reduce the impact of short-term jitter on route switching;

[0085] Under the availability gating, the link cost is obtained by weighted aggregation, and a time proximity penalty is added to the link near the failure window. When in the failure window or disabled by policy, the weight is set to a maximum value to prohibit selection:

[0086]

[0087] wherein, ω e (t) represents the time variation weight (cost value) of link e at time t, dimensionless, the smaller the value, the better, χ e (t) is an availability indicator, taking 1 when in the predicted failure window or disabled, otherwise taking 0, M represents the number of indicators participating in aggregation, a j is the non-negative aggregation weight of the jth index, κ is the failure proximity penalty coefficient, is the start time of the next failure window of link e, Δ is the advance amount, used for cost lifting before failure, Ω is a maximum constant, used to represent the cost of prohibiting routing, and max(·) represents the upper bound operator in the interval;

[0088] ω e (t) is recalculated in a fixed beat or event triggered manner, events include link state change, weather update and load mutation, weight update enters the path search submodule, which is used for multi-objective evaluation function routing and preset switching;

[0089] In this embodiment, the cost time sequence is pushed to the path search submodule in an incremental manner through a structured tool interface, the default upper limit of the number of edges in a single batch transmission is 5000, which can be adjusted to 1000-20000; to limit the calculation overhead, the size of the candidate path set is defaulted to K=4, which can be adjusted to 2-8, the candidate generation timeout is defaulted to 300 ms, which can be adjusted to 100-1000 ms, and the caliber is determined according to the calculation time statistics of the offline reference topology. Optionally, when the incremental update does not cause the optimal path to change and the time margin to the next failure window is greater than twice the advance, the preset refresh is skipped to avoid invalid delivery. In abnormal cases, such as cost update lag or discontinuity, a one-time reevaluation is completed based on the latest complete snapshot and the candidate set is updated.

[0090] Specifically, the above rules unify the geometric, physical and business information into a single time-varying cost, make different dimensions of indicators additive by consistent direction and robust processing, and handle the time sequence risk related to the failure window by gating and proximity penalty; in this way, the routing calculation not only reflects the current performance state, but also reacts to the upcoming disconnection or joint event in advance to avoid packet loss and reordering caused by switching at the critical moment; the definition and calculation process cooperates with the space-time graph model to provide a time function for each edge on the graph, which can be directly used for path search, and can seamlessly connect with the disabled list and aggregation module of the policy verification submodule; the weight update is triggered by both beat and event channels, which takes into account the dynamics and avoids unstable routing caused by frequent fluctuations;

[0091] In one embodiment, the path search submodule calculates a set of candidate paths that meet the policy constraints in the space-time graph model for the data flow to be forwarded, and selects a target path according to a multi-objective routing evaluation function; the target path is preset before the predicted failure link of the data flow arrives and activated when the trigger condition is met;

[0092] Further, the preset adopts the order of installation first and switching later, and the preset advance is consistent with the advance of the aforementioned failure proximity penalty to ensure caliber uniformity; the trigger condition includes at least one of arrival time prediction, link state event and SLA deviation threshold crossing, the upper bound of the error of the arrival time prediction is used to set the trigger debounce time, the debounce is defaulted to 200 ms, which can be adjusted to 50-500 ms; the path stability is guaranteed by a time smoothing window, the window is defaulted to 2 s, which can be adjusted to 0.5-5 s, to reduce frequent switching. If multiple candidate paths are triggered simultaneously, the decision is made in the order of remaining bandwidth, path hop count and distance to failure window.

[0093] The composition of the multi-objective routing evaluation function includes:

[0094] Establishing an evaluation dimension set And the candidate path set is gated by the mandatory constraints of the policy verification submodule, paths violating the disable list or regulatory domain are directly eliminated, and do not enter the scoring;

[0095] The predicted path-level indicators of each candidate path p at time t The sum type uses link summation (such as delay), and the bottleneck type uses minimum or maximum (such as throughput, packet loss), and the statistical caliber is consistent with the network state data;

[0096] After gating, the path score is calculated in the form of weighted sum of structure cost and SLA deviation penalty:

[0097]

[0098] Wherein, J p (t; f) represents the comprehensive score of data flow f on path p at time t, the smaller the value is, the better, λ ω , λ sla is the weight of structure cost and SLA penalty, the value is 0-1, and λ ω + λ sla = 1, E p is the link set contained by path p, |E p | is the cardinality thereof, ω e (t) is the time-varying weight (value) of link e at time t, dimensionless, is the evaluation dimension set, b m (f) is the normalized weight of data flow f on dimension m, non-negative and the sum of each item is 1, max(·) is the maximum value operator, r m is the direction factor, 1 for delay / jitter / packet loss, -1 for throughput / availability, is the path-level prediction value, is the corresponding target value, σ m > 0 is the normalized scale of dimension m, such as delay ms, jitter ms, packet loss and availability are ratios, and throughput is Mbps;

[0099] Time smoothing is performed on J p (t; f) to suppress transient jumps, and the remaining bandwidth is high, the path is shorter, and the failure window is farther;

[0100] The priority vector ρ m (f) is extracted from the high-level task instruction, and is standardized as the evaluation weight:

[0101]

[0102] Wherein, b m (f) is the normalized weight of dimension m, ρ m(f) is the raw priority derived from service level objective resolution, non-negative;

[0103] Weighting profile example: real-time interaction class (voice / control) takes 时延 = 0.40, p 抖动 = 0.30, p 丢包 = 0.15, p 可用性 = 0.10, p 吞吐 = 0.05, bulk transfer class takes 吞吐 = 0.50, p 可用性 = 0.20, p 时延 = 0.15, p 丢包 = 0.10, p 抖动 = 0.05; top-level weight suggestion: real-time interaction (λ sla , λ ω ) = (0.7, 0.3), bulk transfer takes (0.6, 0.4), the above profiles can be fine-tuned according to service playback and offline simulation;

[0104] Select the path with the smallest J p (t; f) in the candidate set as the target path, and form a preset alternative according to the score gap of the suboptimal path for trigger switching;

[0105] Specifically, the evaluation function processes the structural cost of the path and the SLA deviation penalty together, so that the path selection reflects the passability of the link layer and the target achievement of the service layer; the structural term uses the path average of the link cost to suppress the bias of obtaining lower single-link cost by more hops, the penalty term unifies the discrimination direction of different dimensions through the direction factor, and realizes cross-scale comparability through the scale parameter; the weight profile is given by the priority vector resolved from the service level target, and the dimension weight in path scoring is obtained through standardization, which facilitates quick switching of emphasis according to the type of service; the top-level weight is used to adjust the sensitivity between the stability of static structure and the fitting degree of dynamic SLA, so as to maintain the expected routing behavior under different loads and weather conditions; parallel decision and time smoothing supplement the stability requirements of the engineering side, so that the output path does not appear frequent jitter when triggering the preset and switching, facilitating the connection with the policy checking and execution module;

[0106] In one embodiment, the activation includes: completing the pre-installation of the target path before the effective window of the predicted failed link, and performing conditional triggering based on the arrival time prediction, link state change event or threshold alarm; when triggered, performing sequential activation of the old path and the new path to reduce packet loss and reordering;

[0107] Optionally, the sequential activation adopts the semantics of first adding and then deleting: first, a new path is opened at the key node and temporarily forwarded in parallel, then the old path export is revoked, and finally the remaining intermediate node rules are cleaned up; the parallel period is 300 ms by default, and can be adjusted to 100-1000 ms; if it is detected that the reordering exceeds the threshold, the upper limit of the capacity of the buffer and the rearrangement queue is temporarily increased until the switching is completed. In abnormal cases, such as new path installation failure or timeout, immediately fallback to the old path and record the failure fingerprint to avoid repeated attempts on the same target in a short period of time.

[0108] In one embodiment, the rule converter maps the natural language described path intention and policy constraints to a set of instructions for a software defined network controller;

[0109] The instruction set includes: flow table addition, deletion, modification, path intention orchestration, batch transaction, versioning and rollback control, idempotent check and permission audit;

[0110] In this embodiment, the rule converter interfaces with the fixed version of the interface mapping table, and the mapping granularity includes flow table addition, deletion, modification, path intention orchestration, batch transaction, versioning and rollback control, idempotent check and permission audit, etc. The single batch size of the default batch transaction is 200 rules, and can be adjusted to 50-1000. The interface rate limit is 1000 calls per second by default, and can be adjusted to 200-5000. The single transaction timeout is 1s by default, and can be adjusted to 200-5000ms. The idempotent check is based on the rule fingerprint and version number. If the target device returns is inconsistent with the expected, an automatic rollback to the previous stable version is performed. When some devices are rejected, the converter only retries the failed subset and outputs a difference report to avoid repeated delivery to the successfully delivered devices.

[0111] In one embodiment, the network control instruction is used to update the flow table entries of each node along the target path;

[0112] For example, the priority of the flow table entry is configured by default with a more accurate match with a higher priority, and in case of conflict, the priority, creation time and version number are used in order to make a decision; the idle timeout is 30s by default, and can be adjusted to 5-300s; the hard timeout is 600s by default, and can be adjusted to 60-3600s; in order to suppress burst rearrangement, the queue priority is set to be linked with the SLA target, and the real-time interaction class is set to be high priority by default, and the batch transmission class is set to be medium priority by default. If the node resources are insufficient to install all rules, the minimum closed loop rule set related to the current active flow is installed first, and a preset flag is reserved to automatically complete the installation after the resources are released.

[0113] The flow table entry at least includes a matching field and an action field,

[0114] The matching field includes one of a five-tuple, a service identifier or a label identifier;

[0115] The action field includes one or more of a forwarding port, a next hop identifier, a label stack operation, and a queue priority setting;

[0116] In one embodiment, the predicted failed link includes at least one of:

[0117] Adjacent orbital plane inter-satellite links of the cross-seam region, star-ground links with reduced availability due to elevation or weather, temporarily unavailable links due to link congestion or maintenance windows;

[0118] In one embodiment, the policy checking submodule performs constraint consistency checking and simulation verification on the to-be-issued policy, triggers automatic revision or rollback when the checking fails, and performs closed-loop evaluation on the service level target after the policy takes effect and feeds back the evaluation result to the large language model reasoning engine to optimize subsequent decisions;

[0119] In this embodiment, the evaluation takes telemetry and alarms as input, forms two types of indicators, SLA achievement degree and switching quality, according to a sliding window, the window is 60s by default, and can be adjusted to 10-300s; the feedback pace is 10s by default, and can be adjusted to 1-60s; when the evaluation shows that the target path is stable in the long term and the redundant resources are sufficient, the reasoning engine moderately reduces the switching sensitivity parameter without changing the intention to reduce the control plane load; when there are two consecutive substandard or abnormal switches, the automatic revision process of the policy checking submodule is triggered, the minimum revision is limited to the adjustment of the weight aggregation and the scope of the trigger threshold, and the existing disabled list and regulatory boundaries are not changed, so as to ensure that compliance is not destroyed.

[0120] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0121] In addition, those skilled in the art can understand that although some embodiments herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments means that it is within the scope of the present application and forms different embodiments. For example, all the above embodiments can be used in any combination. The information disclosed in the BACKGROUND section is only intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A method for constructing a satellite communication network based on large language model reasoning, characterized in that, include, Step S1: Deploy a satellite network cognitive decision center containing a large language model inference engine on the ground or on the satellite. The decision center interacts with the topology prediction submodule, the path search submodule, and the strategy verification submodule via a structured tool interface. Step S2: Receive multimodal input of network status data and high-level task instructions, wherein the high-level task instructions include service level objectives and constraints; Step S3: The large language model inference engine performs semantic parsing and intent extraction on the multimodal input to generate a network configuration strategy framework that includes path intent, resource and policy constraints, and target SLA. Step S4: The topology prediction submodule is called to generate a spatiotemporal graph model for the target time period, and the consistency and security are verified by the strategy verification submodule. Step S5: The policy framework is mapped to the application programming interface call instructions of the software-defined network controller through the rule converter; Step S6: The software-defined network controller sends network control commands to the relevant nodes of the satellite communication network to update flow table entries in order to construct and activate the target path.

2. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The network status data includes at least one of the following: satellite orbit ephemeris data, inter-satellite link status data, load and availability data of ground gateway stations and border gateways, real-time or forecast meteorological data, and historical routing and alarm logs.

3. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The large language model inference engine parses high-level task instructions, including: Extract task objectives, target areas, service level objectives, and mandatory constraints, and bind the extraction results into machine-readable policy constraints.

4. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The topology prediction submodule is constructed using a spatiotemporal graph model with satellites and ground gateway stations as nodes, inter-satellite links and satellite-to-ground links as edges, and the link weights changing over time. It also marks the predicted failed links and their effective time windows. Link failures include interruptions in the cross-joint area caused by the relative motion of satellites. The time function for the link weight is defined as follows: Based on ephemeris and station data, the visibility, slant range and elevation angle of the link are obtained, and the predicted failure time window of the cross joint is generated. The list of disabled items and the maintenance window are merged into the same availability time axis. By combining frequency band and power link budget, weather attenuation and link load, a raw sequence of indicators such as latency, jitter, packet loss and available bandwidth is obtained, and a representative value at the current moment is extracted using a sliding window. Each original index is mapped to a cost-type quantity in the range [0,1] and then pruned to upper and lower limits. The mapping process is as follows: When the direction of a certain indicator being worse or better is inconsistent, it is first processed with a directional factor, and then linear segmentation is performed to make indicators of different dimensions comparable. in, y represents the j-th normalized cost of link e at time t, taking values ​​[0,1]. e,j (t) represents the corresponding original index, L j U j These represent the lower and upper limits of the engineering scope for this indicator, respectively. j This is a direction factor; a value of 1 indicates that the larger the value, the greater the cost, and a value of -1 indicates that the smaller the value, the greater the cost. right Outlier suppression and exponential smoothing are applied, and the time constant is set according to the timeliness of the business to reduce the impact of short-term jitter on routing switching; Under availability gating, link costs are obtained through weighted aggregation, and time proximity penalties are added to links nearing failure windows. When a link is in a failure window or disabled by a policy, its weight is reset to its maximum value to prevent it from being selected. Where, ω e (t) represents the time-varying weight of link e at time t, which is dimensionless and a smaller value is better. χ e (t) represents the availability indicator, set to 1 if the indicator is in the predicted failure window or disabled, and 0 otherwise. M represents the number of metrics involved in the aggregation. j Let κ be the non-negative aggregate weight of the j-th index, and let κ be the failure proximity penalty coefficient. Let be the start time of the next failure window for link e, Δ be the lead time used to increase the cost before failure, Ω be a maximum constant used to represent the cost of prohibiting route selection, and max(·) be the operator for taking the supremum of the interval. Recalculate ω using a fixed beat or event-triggered method e (t), events include link status changes, weather updates and load mutations. Weight updates are entered into the path search submodule for use by multi-objective evaluation functions for route selection and preset switching.

5. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The path search submodule in the spatiotemporal graph model calculates a set of alternative paths that satisfy the policy constraints for the data stream to be forwarded, and selects the target path according to the multi-objective routing evaluation function. The target path is preset before the data stream reaches the predicted failed link and is activated when the triggering condition is met. The multi-objective routing evaluation function is composed of: Establish a set of evaluation dimensions The candidate path set is gating by the mandatory constraints of the strategy verification submodule. Paths that violate the prohibited list or regulatory domain are directly eliminated and do not enter the scoring process. For each candidate path p, a predicted path-level index is obtained at time t. For summation categories, link summation is used; for bottleneck categories, minimum or maximum value is used, and the statistical caliber is consistent with the network status data. After passing the gate, the path score is calculated as a weighted average of structural cost and SLA deviation penalty: Among them, J p (t; f) represents the comprehensive score of data stream f at time t via path p; the smaller the value, the better. λ ω ,λ sla Let λ be the weight of the structural cost and the SLA penalty, taking values ​​between 0 and 1, and λ be the weight of the SLA penalty. ω +λ sla =1,E p Let |E be the set of links contained in path p. p | is its base, and the unit is bars, ω e (t) represents the time-varying weight of link e at time t, which is dimensionless. b is the set of evaluation dimensions m (f) represents the normalized weights of the data flow f in dimension m, which are non-negative and sum to 1. max(·) is the maximum value operator. m As a direction factor, latency / jitter / packet loss are set to 1, and throughput / availability are set to -1. These are path-level predicted values. For the corresponding target value, σ m >0 represents the normalized scale for dimension m; For J p (t;f) performs time smoothing to suppress instantaneous jumps, and when parallel, it prioritizes the one with higher remaining bandwidth, shorter path and farther failure window; Extract the priority vector ρ from the high-level task instructions. m (f), and standardized into evaluation weights: Among them, b m (f) represents the normalized weights of dimension m, ρ m (f) represents the original priority obtained from the service level objective resolution, which is non-negative; Select J from the candidate set p The path with the smallest (t;f) is taken as the target path, and the second-best paths are formed into preset alternatives according to the score difference, which are used for trigger-based switching.

6. The satellite communication network construction method based on large language model reasoning as described in claim 5, characterized in that, The activation includes: pre-installing the destination path before the effective window of the predicted failed link, and triggering the condition based on arrival time prediction, link status change event or threshold alarm. When triggered, the old path and the new path are sequentially activated to reduce packet loss and reordering.

7. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The rule converter maps the path intent and policy constraints described in natural language into a set of instructions for software-defined network controllers. The instruction set includes: adding, deleting, and modifying flow tables; path intent orchestration; batch transaction distribution; versioning and rollback control; idempotency verification; and permission auditing.

8. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The network control instructions are used to update the flow table entries of each node along the target path; The flow table entry includes at least a matching field and an action field. Matching fields include one of the following: a 5-tuple, a service identifier, or a tag identifier; The action field includes one or more of the following: forwarding port, next hop identifier, tag stack operation, and queue priority setting.

9. The satellite communication network construction method based on large language model reasoning as described in claim 4, characterized in that, The predicted failure link includes at least one of the following: Inter-satellite links between adjacent orbital planes in the cross-joint area, satellite-to-ground links whose availability is reduced due to elevation angle or weather, and links that are temporarily unavailable due to link congestion or maintenance windows.

10. The satellite communication network construction method based on large language model reasoning as described in claim 1, characterized in that, The strategy verification submodule performs constraint consistency verification and simulation verification on the proposed strategy. If the verification fails, it triggers automatic revision or rollback. After the strategy takes effect, it performs a closed-loop evaluation of the service level target through telemetry reports and feeds the evaluation results back to the large language model inference engine to optimize subsequent decisions.

Images