Low earth orbit constellation emergency data transmission task planning method, system and device

Abstract

The application provides a low-orbit constellation emergency data transmission task planning method, system and device, a mathematical model of a low-orbit constellation emergency data network transmission task planning problem is established based on a resource-restricted space-time evolution diagram, and a network dynamic topology, a task and a space-time coupling relationship of resources are accurately described; on this basis, a two-stage optimization solution enabled by machine learning is proposed: a link importance predictor is trained by using a data-driven method in a link establishment scheme generation stage, and link establishment decisions are guided; in a transmission scheme optimization stage, a data transmission planner is constructed by using a linear programming method to complete accurate planning of data flow. A large number of simulation scene experiments show that, compared with a classic dynamic optimization method such as a rolling time domain optimization, the total task income of the scheme is increased by more than 25%, and the scheme can significantly improve the data transmission efficiency, and provides a theoretical basis and an engineering reference for large-capacity emergency data network transmission of a low-orbit constellation.

Description

Technical Field

[0001] This invention relates to the fields of inter-satellite data transmission, inter-satellite network structure, and emergency data transmission mission planning, and in particular to a transmission mission planning method, system, and device for networked transmission of emergency data from low-Earth orbit constellations based on machine learning. Background Technology

[0002] In recent years, low-Earth orbit (LEO) satellites have been widely used in geological exploration, disaster early warning, and meteorological monitoring due to their advantages of low cost, wide coverage, and high flexibility. In future multi-satellite autonomous planning scenarios, LEO satellites can autonomously perceive mission requirements, plan and execute imaging tasks through onboard computing units. Afterwards, onboard preprocessing can quickly assess the urgency of the mission, distinguishing between time-sensitive emergency tasks (such as disaster early warning) and routine tasks with lower time requirements, and finally transmitting the mission data back. For high-precision data initially determined to be "emergency" onboard, to ensure timeliness, a real-time transmission plan needs to be planned to quickly transmit the data back to the ground center for in-depth analysis, effectively avoiding the risk of decision-making delays caused by transmission delays. However, the traditional backhaul mechanism relying on transit windows is insufficient to meet the real-time transmission needs of large-capacity, high-precision emergency data.

[0003] Against this backdrop, the problem of high-capacity emergency data network transmission for low-Earth orbit constellations (LLHVEDTP) has emerged. This problem focuses on the data transmission and scheduling stage after satellites acquire raw mission data. Its core challenges stem from three coupled constraints: First, the network topology is dynamically time-varying, as satellite movement along its orbit causes continuous changes in the inter-satellite visibility window, resulting in periodic fluctuations in link connectivity; second, onboard resources are strictly limited, as low-Earth orbit satellites are constrained by size and payload, resulting in extremely limited onboard buffer capacity and energy supply capabilities, further restricting the flexibility of transmission scheduling; and third, emergency missions arrive dynamically, and the suddenness and unpredictability of emergency missions place stringent demands on the system's online decision-making capabilities and rapid response efficiency.

[0004] From a problem structure perspective, LLHVEDTP comprises two coupled phases: link scheduling and data flow scheduling. The link scheduling phase is essentially a dynamic link scheduling problem with resource constraints. LLHVEDTP not only requires planning the link establishment scheme between satellites and ground stations but also needs to consider link establishment decisions between satellites, adding on-board resource constraints and mission dynamic constraints. After determining the link establishment scheme, data flow scheduling needs further optimization to achieve efficient resource allocation. Therefore, the constraint coupling and dynamism of LLHVEDTP are more prominent, significantly increasing the problem complexity.

[0005] Systematic investigation revealed that the LLHVEDTP problem remains unresolved. Regarding traditional methods, precise solution algorithms and heuristic methods based on conflict graphs, applicable to small satellite connectivity design, effectively address the task-aware problem of small, resource-constrained satellite connectivity design. Benders decomposition solves the maximum flow problem in small satellite networks. Contact graph routing, a classic routing scheme for delay-tolerant networks, has been validated in deep space networks, but suffers from undetected contact failures, potentially leading to failed links in transmission paths and reduced packet delivery rates. Knowledge-based hybrid evolutionary algorithms improve connectivity accuracy, but their simulation scenarios are still based on fixed topology assumptions and do not consider dynamic link switching caused by satellite orbital motion. To make full use of multiple low-Earth orbit satellites with limited communication opportunities, existing methods have proposed approaches such as prioritizing satellites with insufficient visible time windows to offload large amounts of data, or using inter-satellite links to offload data between satellites before establishing connections with ground stations, or prioritizing maximizing the visible time windows between low-Earth orbit satellites and ground stations to increase the amount of data transmitted when downloading data to multiple ground stations. However, the above studies do not consider the latency and integrity of data transmission and are difficult to handle mission planning in emergency scenarios, resulting in insufficient timeliness.

[0006] Furthermore, regarding response speed for time-sensitive tasks, existing technologies suffer from several drawbacks. First, they fail to consider onboard energy constraints—low-Earth orbit satellites have limited energy, and excessive pursuit of timeliness could lead to insufficient energy for subsequent missions. Second, they do not adequately consider inter-satellite multi-hop scenarios and the segmented transmission dependencies of large-capacity data, making them susceptible to overall transmission failure due to single-segment loss. Third, the algorithms employed in these existing technologies are highly complex, insufficiently consider the "extreme dynamics" and "engineering constraints" of LLHVEDTP, and have shortcomings in adaptability to large-scale dynamic scenarios and algorithm convergence speed, resulting in insufficient efficiency and accuracy. Fourth, existing solutions combining machine learning generally fail to consider complex business characteristics, do not assess onboard computational overhead, and are highly dependent on training data. They often focus on optimizing single aspects (such as routing or offloading), leading to high training costs and long training cycles, and lacking a solution framework suitable for LLHVEDTP optimization requirements.

[0007] In the LLHVEDTP problem, the challenge of decision-making lies in how to optimize decisions based on real-time task information when future task information is unknown, while also taking into account the potential benefits of future tasks, avoiding resource waste due to short-sighted decisions, and improving decision robustness. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, and device for planning emergency data transmission missions for low-Earth orbit constellations, thereby solving at least one technical problem existing in the prior art. Specifically, this invention provides the following technical solutions: In a first aspect, the present invention provides a method for planning emergency data transmission missions for low-Earth orbit constellations, the method comprising: S1. Embed constellation resource status as node attributes and describe topological connectivity and resource availability simultaneously to construct a spatiotemporal evolution graph model. S2. Determine the decision variables and objective function for the task planning problem, and define the constraints. S3. Construct a two-stage optimization model and a training dataset. Train the two-stage optimization model using supervised learning to obtain the trained two-stage optimization model. S4. Based on the trained two-stage optimization model, plan the emergency data transmission mission for the low-Earth orbit constellation.

[0009] Furthermore, in S1, the specific method for constructing the spatiotemporal evolution diagram model is as follows: First, the overall planning duration is broken down into T consecutive time slots. Each time slot t corresponds to a static subgraph. , where the set of nodes Includes satellite nodes With ground station nodes , The total number of satellites in the constellation. The total number of ground stations, satisfying , Represents the set of potential links. Indicates resource attributes; Secondly, in time slots Inside, , and All remain unchanged, and this time slot is regarded as a static network slice, within which link selection and data allocation decisions are made; Furthermore, when transitioning from time slot t to t+1, the satellite's orbital motion causes... The relative position changes of satellite nodes in the middle, which in turn trigger The update, and the data transmission within time slot t, will consume satellite buffer resources, making... based on The amount of data transmitted is dynamically updated. Furthermore, potential links between satellites exist only within the visible window, and active links are filtered through a connection plan; the connection plan is defined as a set of conflict-free connections after scheduling, which is a subset of potential links.

[0010] Furthermore, the potential connection set Includes a subset of inter-satellite links With satellite-to-ground link subset ,satisfy Resource attributes Defined as the resource state vector of a satellite node, for a satellite Its resource attributes are represented as ,in For satellite The remaining cache capacity in time slot t, For satellite The remaining energy.

[0011] Furthermore, in S2, the decision variables include: Link connectivity variables represented by binary integer variables , indicating the connectivity status of link e within time slot t; Data transfer variables This represents the amount of data transmitted in task f within link e in time slot t. , The set of data transmission tasks arriving in time slot t; Task completion rate represented by a binary integer variable , indicating time slot The degree of completion of task f at the end.

[0012] Furthermore, in S2, the optimization objective is to maximize the value of the complete mission data transmitted back to the ground station up to the current time slot t. Therefore, the specific objective function is:

[0013] in, The value of task f; The completion rate of task f, This is the set of data transmission tasks arriving in time slot t.

[0014] Furthermore, in S2, the constraints include: connection quantity constraints, data flow constraints, cache constraints, energy constraints, and task integrity constraints; The data flow constraints include link capacity limit constraints, task data arrival constraints, traffic conservation constraints, and ground station aggregation constraints. The energy constraints include solar charging constraints and energy balance constraints after determining the energy consumption for data forwarding, data receiving, and basic operation.

[0015] Furthermore, in S3, the two-stage optimization model includes an input module, a chain building scheme generation module, and an optimization module; The input module is used to determine various input characteristics in the system state and to determine the decision evolution rules; The chain building scheme generation module is used to map the system state to the link value through supervised learning, and obtain a set of feasible links; The optimization module optimizes the set of feasible links using linear programming to select the optimal link plan.

[0016] Furthermore, the input features include: Satellite status characteristics N represents the number of satellites, and K represents the number of missions; mission status characteristics Network state characteristics M represents the number of ground stations, and the element This indicates that node i (satellite / ground station) and node j are visible in time slot t and can establish a link; otherwise, it is 0. The above features together constitute the state feature vector: .

[0017] Furthermore, the execution flow of the chain establishment scheme generation module is as follows: Receive the state feature vector sent by the input module Output the predicted value of each link. , where e is the link number; This module is based on Filter links from high to low to obtain a set of feasible links. , as input to the optimization module; In this module, the link value is set as follows: First, in a static scenario, the contribution value of each link is obtained through a two-stage heuristic method, which serves as the label for the value of each link in a dynamic scenario; in a dynamic scenario, through model training and prediction, the maximum task value increment that can be obtained in the future after selecting a certain link in the current state is taken as the link value.

[0018] Furthermore, the optimization module optimizes the set of feasible links using linear programming, that is, the objective function in S2 is set as:

[0019] in, The disturbance is a very small positive number.

[0020] Furthermore, in S3, training the two-stage optimization model first requires constructing a training dataset, which is done as follows: First, initialize the training dataset. The simulation generates multiple static cases, forming a static case set. All static examples have known full-time-slot task arrival and topology changes; Then, for each instance in the instance set, the full-time-slot task requirements for that instance are extracted. With network topology The example is then decomposed into T time-slot subproblems. Each sub-problem corresponds to a time slot state in a dynamic scenario. ; Secondly, for each time slot subproblem Extracting state feature vectors The solution is obtained using a two-stage temporal heuristic method, which extracts the link selection scheme. The true value of the corresponding link Forming state-label pairs ; Finally, the result will be Add D and return dataset D; dataset D is used as the training dataset and is divided into training and validation sets for model training.

[0021] Furthermore, in S3, the model training method is as follows: First, the training dataset is randomly divided into a training set and a validation set according to a certain ratio; Then initialize the training parameters, including setting the learning rate, convergence threshold, and iteration counter; Iterative training in a loop: If the iteration counter is less than the preset number of training epochs and the change in validation loss is greater than the convergence threshold, repeat the following operation: Shuffle the training set. And divide the samples according to batch size Batch training: For each batch Extract sample status within the batch With tags Then predict through the model Then calculate the loss. The model parameters are updated by backpropagation based on the loss calculation results. Calculate the current validation loss using the validation set. The iteration counter is incremented by 1; where, For learning rate, This represents the parameter change in a two-stage optimization model. Represents the validation set. These represent the parameters of the two-stage optimization model; After the loop terminates, output the trained model. .

[0022] Secondly, the present invention also provides a low-Earth orbit constellation emergency data transmission mission planning system, the system comprising: The spatiotemporal evolution graph construction module is used to embed constellation resource status as node attributes and synchronously describe topological connectivity and resource availability in order to construct a spatiotemporal evolution graph model. The initial setup module is used to determine the decision variables and objective function of the task planning problem, and to define the constraints. The two-stage optimization model module constructs and stores the trained two-stage optimization model, and plans the low-Earth orbit constellation emergency data transmission mission based on the trained two-stage optimization model. The model training module is used to construct the training dataset and train the two-stage optimization model through supervised learning.

[0023] Thirdly, the present invention also provides a low-Earth orbit constellation emergency data transmission mission planning device, which includes at least a storage device and a processor. The processor calls data and computer instructions in the storage device to execute the low-Earth orbit constellation emergency data transmission mission planning method as described above.

[0024] By adopting the above technical solution, the present invention has the following beneficial effects: To address challenges such as dynamic topology, random task arrival, and resource constraints, this solution constructs a mathematical model of the problem and proposes a two-stage optimization method (MLETOM) empowered by machine learning, integrating supervised learning and linear programming. This solution guides the generation of link establishment and data transmission schemes by predicting link importance, effectively reducing the impact of uncertainty on scheduling performance. Compared to existing solutions, this solution exhibits better robustness and generalization. It can effectively learn strategies from optimal solutions, ensuring complete transmission of the corresponding tasks in each newly arriving task set, thus providing a more stable and high-quality solution. Furthermore, through the synergy of "link value prediction—optimized data allocation," this solution avoids the performance degradation caused by resource exhaustion or link selection errors in traditional algorithms, better adapting to changing task requirements in dynamic environments, thereby achieving higher total task benefits.

[0025] Compared to the STPHA method, this approach employs a different strategy in the link selection phase. STPHA relies on manually designed heuristic rules, evaluating links through empirical importance metrics, and is only applicable to static scenarios—that is, information such as task input and arrival time is fully known before optimization. In contrast, this approach uses a data-driven method to intelligently evaluate link value, making it better suited to complex and dynamic environments.

[0026] In the subsequent linear programming phase, STPHA introduces a forward-looking objective function, which enhances the robustness of the solution by weighting the visibility information of multiple future time slots. This solution, however, is designed for online optimization scenarios where system states arrive sequentially and future information is unknown. Therefore, it makes decisions based solely on the current state, ensuring real-time performance while adapting to dynamically changing network environments.

[0027] Compared with classical methods, the MLETOM method improves the overall task reward by over 25%, validating its effectiveness and superiority in link optimization and data transmission planning. The problem studied in this proposal exhibits typical dynamic resource constraints, and the proposed MLETOM method provides a scalable solution approach for similar network optimization problems. Attached Figure Description

[0028] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a spatiotemporal evolution graph model with resource constraints, as described in an embodiment of the present invention. Figure 2 A schematic diagram of a two-stage optimization method for machine learning empowerment in an embodiment of the present invention; Figure 3 This is a schematic diagram of the link selector operation according to an embodiment of the present invention; Figure 4 Box plots showing the distribution of total revenue for different algorithm tasks; Figure 5 A comparison chart showing the incremental revenue of each comparison algorithm in each time slot; Figure 6 A comparison of total task benefits over time; Figure 7 This is a schematic diagram of the main method flow of an embodiment of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Those skilled in the art should understand that the following specific embodiments or implementation methods are a series of optimized configurations listed to further explain the specific content of the invention. These configuration methods can be combined or used in conjunction with each other, unless the invention explicitly states that some or a specific embodiment or implementation method cannot be associated with or used in conjunction with other embodiments or implementation methods. Furthermore, the following specific embodiments or implementation methods are merely optimized configurations and are not intended to limit the scope of protection of the invention.

[0032] The present invention will be further explained and described below with reference to specific embodiments.

[0033] This proposal focuses on the networked transmission of high-capacity emergency mission data in Low Earth Orbit (LEO) constellations. Its core objective is to achieve efficient and reliable transmission of real-time emergency data to ground stations under resource constraints through coordinated planning of inter-satellite and satellite-to-ground links. The main research object of this proposal is the LEO constellation system and the emergency data missions it needs to process.

[0034] The research object addressed in this scheme mainly involves inter-satellite links (S2S) and satellite-to-ground links (S2G). Inter-satellite links refer to high-speed data transmission links between low-Earth orbit satellites, serving as the core carrier for realizing multi-hop relay transmission of large-capacity data. Their connectivity is determined by the "visibility window" caused by satellite orbital motion, exhibiting dynamic and time-varying characteristics. Due to resource constraints, a single satellite can only establish a communication link with a maximum of one satellite. Satellite-to-ground links refer to data transmission links between low-Earth orbit satellites and ground stations. Constrained by the geographical location of the ground station and the satellite's transit time, to avoid link conflicts, a single ground station can only establish a connection with a single satellite at any given time.

[0035] In this embodiment, the main technical problem addressed by this solution can be summarized as follows: Within any planning period, given the current arrival status of emergency tasks (number of tasks, amount of data, priority), topology (set of available links for S2S and S2G), and satellite resource status (remaining buffer capacity), the following decisions need to be made: 1. Link selection: Select conflict-free active links from available S2S and S2G links to construct a "link-resource" matching set that includes time-series characteristics; 2. Data scheduling: Allocate emergency task data to selected links to ensure that it is gradually delivered to the ground station through multi-hop transmission in subsequent time slots, while satisfying constraints such as no satellite buffer overflow and priority transmission for high-priority tasks. Each emergency task data needs to adapt to link resources in the dynamic network topology to achieve networked data transmission.

[0036] The dynamic topology and satellite resource constraints of low-Earth orbit constellations are the main challenges of LLHVEDTP, which is often characterized by Time-Evolving Graphs (TEGs). However, traditional TEG models can only depict the time-varying characteristics of the topology and cannot reflect the dynamic consumption and transfer processes of onboard resources (such as buffers). Combining... Figure 7 The implementation process of this embodiment mainly includes the following aspects: I. Establishment of a Spatiotemporal Evolution Diagram Model To quantitatively characterize the coupling relationship between "topology time-varying—resource constraints—task dynamics," this scheme, based on TEG, introduces an on-board resource transfer mechanism to construct a resource-constrained spatiotemporal evolution graph model (STEG-R), achieving unified modeling of network topology and resource status, such as... Figure 1 As shown, resource status is embedded as a node attribute into the spatiotemporal graph to achieve a synchronous description of topological connectivity and resource availability. First, the overall planning duration is decomposed into T consecutive time slots. Each time slot t corresponds to a static subgraph. , where the set of nodes Includes satellite nodes With ground station nodes N is the total number of satellites in the constellation, and M is the total number of ground stations, satisfying the following conditions: According to time slots Visibility state, set of potential connections Includes a subset of inter-satellite links (Saving inter-satellite visibility) and satellite-to-ground link subset (Satisfies satellite and ground station visibility requirements), satisfies and Resource attributes Defined as the resource state vector of a satellite node ,in For satellite The remaining cache capacity in time slot t, For satellite nodes The remaining energy (electricity).

[0037] The STEG-R model in this embodiment achieves collaborative characterization of topology and resources through a "static description within time slots - dynamic transfer between time slots" approach: within time slot t, , and If all remain unchanged, this time slot can be considered a "static network slice," within which link selection and data allocation decisions are made, such as distributing task packets to... The link in the middle, while ensuring that the data volume does not exceed the satellite's When transitioning from time slot t to t+1, the satellite's orbital motion causes... The relative position changes of satellite nodes in the middle, which in turn trigger The update of (link connectivity) and the data transmission within time slot t will consume satellite buffer resources, making... based on With the amount of data transmitted dynamically updated, such as by satellite If x units of data are received in time slot t, then Furthermore, the STEG-R model is compatible with the latency-tolerant network characteristics of low-Earth orbit constellations: potential links between satellites only exist within the visible window, and some potential links cannot be activated due to resource conflicts (such as insufficient satellite cache, link bandwidth occupation, etc.). Therefore, the model needs to filter active links through a Contact Plan (CP). A CP is defined as a set of conflict-free connections after scheduling, which is a subset of potential links and is the input for subsequent data transmission optimization.

[0038] II. Mathematical Setting of the Problem Building upon the STEG-R method described in this embodiment, we further elucidate the mathematical expression of the LLHVEDTP problem. In this embodiment, the resource allocation problem under dynamic network topology is transformed into the Extended Multi-Commodity Flow Problem (MCFP), and a Mixed Integer Programming (MIP) model is constructed. The specific details are as follows: 1. Decision variables mainly include the following: Link connectivity variables: using binary integer variables This indicates the connectivity status of link e within time slot t. , Let be the set of links in time slot t in the STEG-R model. If Then link e is activated in time slot t; if If so, then link e is not activated.

[0039] Data transfer variable: This variable is defined as a continuous variable. This represents the amount of data transmitted in task f within link e in time slot t. , This is the set of data transmission tasks arriving in time slot t.

[0040] Task integrity variable: This variable is defined as a binary integer variable. This indicates the degree of completion of task f at the end of time slot t. If... Then all data for task f Transmitted to ground station; if If so, then task f is not completed.

[0041] 2. Set the objective function In this embodiment, the optimization objective is to maximize the value of the complete mission data transmitted back to the ground station up to the current time slot t: (1) In the formula: The value of a task set for the Mission Control Center (MMC); This is a completion variable for task f, i.e., a task integrity variable, which contributes value only when all data of task f is transmitted to the ground station. .

[0042] 3. Set constraints Mission data is transmitted in the form of data streams between low-Earth orbit satellites in the satellite network. This embodiment also considers the constraints of the number of connections, data streams, energy, and mission integrity.

[0043] (1) Connection quantity constraint Satellite connection number constraint: Each satellite in a time slot A maximum of one inter-satellite link (S2S) and one satellite-to-ground link (S2G) can be activated within the system to avoid resource conflicts.

[0044] In the formula: These are the inter-satellite and satellite-to-ground link subsets for time slot t, respectively.

[0045] Ground station connection limit: Each ground station can activate a maximum of 1 satellite-to-ground link within time slot t to avoid signal interference.

[0046] (2) Data flow constraints Link capacity limit constraint: The total amount of all task data transmitted on link e within time slot t must not exceed the product of the maximum transmission capacity of link e and the link's active state. (5) In the formula: This represents the maximum transmission capacity of link e in time slot t. Link activation variable; inactive links The data transfer volume is forced to 0.

[0047] Task data arrival constraint: The total amount of data transmitted by task f within time slot t must not exceed the sum of its available data volume and its cached data volume.

[0048] in, In the formula: the second term on the right is the net cache accumulation of task f in the first t-1 time slots. This refers to the link flowing into the system, i.e., the inter-satellite link. This indicates the link outgoing from the system, i.e., the satellite-to-ground link that transmits data back to the ground station. Indicates time slot Network topology diagram, Indicates time slot Time through link The amount of data flowing into the system for task f. Indicates time slot Time Link The amount of data transmitted back to the ground station for task f.

[0049] Flow conservation constraint: For any satellite node i, time slot The difference between the amount of data received for mission f and the amount of data transmitted is equal to the change in the satellite's buffer size for mission f.

[0050] in, In the formula: For satellite i in time slot t, the mission Cache; This buffers the data from the previous time slot, ensuring a balance in the "receive-transmit-store" process at the satellite nodes.

[0051] Ground station convergence constraints: Mission f completed If and only if all data is transmitted to the ground station:

[0052] in, In the formula: The source satellite for mission f; This is a set of ground stations. The left side represents the total amount of data transmitted to the ground stations by the time slot t of mission f, and the right side represents the total amount of data in the completed state. , .

[0053] (3) Cache constraints

[0054] In the formula: This represents the maximum buffer capacity of satellite i; the left side shows the time slot of satellite i. The total cache usage for all tasks must be strictly controlled within the capacity limit.

[0055] (4) Energy confinement Data forwarding energy consumption: The energy consumption of satellite i forwarding data within time slot t is directly proportional to the amount of data forwarded, the forwarding power, and the time slot duration.

[0056] In the formula: Let e ​​be the transmission rate of link e; To ensure that data transmission duration aligns with the physical meaning of "power × time," This represents the amount of data transmitted by link e for task f in time slot t. Indicates the time slot length.

[0057] Data reception energy consumption: satellite within time slot t Energy consumption for receiving data It is directly proportional to the amount of data received, the receiving power, and the duration of the time slot:

[0058] In the formula: This indicates the power used when receiving data.

[0059] Basic operating energy consumption: Energy consumption for satellite i to maintain basic operation within time slot t. , is the product of a fixed value and the time slot duration:

[0060] In the formula: This is the basic operating power for the satellite.

[0061] Solar charging constraint: Solar charging amount of satellite i within time slot t It is directly proportional to the maximum solar power and the effective sunshine time within the time slot:

[0062] In the formula: Let t be the shadowing time of satellite i in time slot t.

[0063] Energy balance constraints: time slots Last satellite Remaining energy It equals the remaining energy in the previous time slot plus the charging capacity minus the total energy consumption, and does not exceed the product of the maximum energy and the safety factor:

[0064]

[0065] In the formula: Let be the remaining energy of satellite i in time slot t; Energy safety factor; This represents the maximum energy of satellite i, ensuring that the satellite's energy does not overflow and that a safe margin is maintained.

[0066] (5) Task integrity constraints

[0067] in, In the formula: the right side is the transmission completion ratio of task f up to time slot t. It can only be 1 when the ratio reaches 1, so as to strictly ensure data integrity.

[0068] III. Two-stage optimization based on machine learning Combination Figure 2As shown, the core process of the LLHVEDTP problem includes "link establishment" and "data transmission," and the rationality of the link establishment plan directly determines the data transmission efficiency. The key to link establishment decision-making lies in accurately predicting the link value. Based on this, this solution proposes a two-stage optimization scheme based on machine learning, which improves the robustness and efficiency of emergency data transmission through the synergy of dynamic link value prediction and precise optimization.

[0069] The overall framework of the solution, namely the two-stage optimization model (abbreviated as MLETOM model), is based on the following: transferring the optimal link value pattern in static scenarios to dynamic scenarios, constructing a link value prediction model through supervised learning, and combining linear programming to achieve coordinated optimization of dynamic link establishment and data transmission scheduling. This is divided into a prediction stage and an optimization stage, with the specific process as follows: Figure 2 As shown, the main steps include the following: (1) State Representation: Integrate network state, satellite state, mission state and time characteristics to construct a high-dimensional comprehensive state vector, providing input for link value prediction. (2) Link Value Prediction: Train a supervised learning model based on historical data, input the comprehensive state vector, and output the value score of each potential link to achieve accurate prediction of dynamic link value. (3) Link Selection and Data Transmission Scheduling: Select the optimal link set based on the predicted link value, use it as a constraint input to the linear programming model, optimize the data transmission allocation scheme, and ensure that the constraints of satellite buffer, energy and link capacity are met. (4) Scheduling Scheme Generation: Output the link establishment plan and data transmission scheduling scheme for the current time slot.

[0070] Next, we will elaborate on how the model was built.

[0071] 1. Dynamic system modeling To clarify the inputs, outputs, and decision boundaries of the two-stage optimization model framework, the problem is remodeled as a dynamic system, mainly including state, action, feasible solutions, system evolution, strategy, and optimization objective.

[0072] (1) System status The system state is updated dynamically over time slots and needs to reflect both historical accumulated information and current sudden changes, defined as a vector. The meanings of each component are shown in Table 1: Satellite status characteristics N represents the number of satellites, K represents the number of missions, and the satellite status characteristics include the remaining energy of a single satellite (percentage of maximum capacity), cache utilization rate (used / total capacity), and the amount of cached data for each mission.

[0073] Task status characteristics : Includes the source satellite, remaining data volume, and value weight (priority) for a single mission, where K represents the number of missions.

[0074] Network state characteristics M represents the number of ground stations, and the element is... This indicates that node i (satellite / ground station) and node j are visible in time slot t and can establish a link; otherwise, it is 0.

[0075] Table 1 Input Features of Link Value Predictor

[0076] (2) Optimization strategy A deterministic optimization strategy can be expressed as:

[0077] The input is the current state. The core of the two-stage optimization model framework is to output feasible decision solutions through "prediction + optimization". The approximate optimality is found. Specifically, within the planning time slot, the deterministic strategy... This determines the value of each potential link and provides a link building plan. To guide data transmission planning The objective function is formulated to determine the task rewards to be completed by the deadline slot, which is the optimization objective function described in Part 1 above.

[0078] (3) Feasible decision-making and system evolution A feasible solution refers to a link establishment scheme and data transmission plan that satisfies all the above constraints within a given time slot. System evolution is mainly determined by the execution of feasible decision schemes and changes in the system's dynamic environment.

[0079] System status from arrive The evolution is driven by two parts: Environment Update: New Missions Arrive Changes in satellite orbital motion (Topological adjacency matrix).

[0080] Decision Execution: Link Selection With data transmission This leads to updates to satellite energy and cache status. The amount of remaining data for the task has decreased. .

[0081] 2. Link Establishment Scheme Generation - Link Selector The generation of the solution, also known as the prediction stage, is essentially about building a supervised learning model to achieve the mapping from "system state to link value".

[0082] (1) Definition of Link Value First, in a static scenario, the temporal two-stage heuristic method (STPHA) obtains the contribution value of each link through global optimization—that is, the amount by which the link improves the overall task value—and uses this as a label for the link value in a dynamic scenario. In a dynamic scenario, the link value is defined as "the maximum task value increment that can be obtained in the future after selecting this link in the current state," which needs to be approximated by model prediction.

[0083] (2) Working Logic The link value predictor, also known as the prediction layer, is illustrated in the diagram below. Figure 3 As shown, the specific process is as follows: Input comprehensive state feature vector (Including satellite, mission, network, and time characteristics).

[0084] The model outputs the predicted value of each link. (e is the link number).

[0085] Link selection: based on Links are filtered from high to low, and a set of feasible links is obtained by combining the "single-satellite-single-station connection constraint". , as input to the optimization layer.

[0086] Preferably, the statistical model can be implemented using a graph neural network model from the prior art.

[0087] 3. Transmission scheme optimization phase The optimization phase can be achieved through an optimization layer, which uses the set of feasible links output by the prediction layer. With this as input, the original mixed-integer programming problem is transformed into a linear programming model, optimizing the task data allocation for each link. This ensures that resource constraints are met and task value is maximized. Specifically, given a selected set of links... In this embodiment, the data transmission problem up to time slot t is modeled as a linear programming problem:

[0088] The constraints given by equations (5) to (16) are satisfied.

[0089] In the formula: To prevent the objective function from being zero, a very small positive number is added. As a perturbation, it is only used for numerical stability and does not affect the optimization results.

[0090] 4. Model Training (1) Forward-looking decision-making training set The decision-making difficulty in dynamic scenarios stems from the "unknown future task / topology," while the optimal solution of the two-stage temporal heuristic method for static scenarios (with known future information) can provide ideal decision labels. First, a prospective training dataset is formed, which is done as follows: Input: Static test case set I, total time slots T, number of nodes N+M; Output: Training dataset The process is as follows: First, initialize the training dataset. Multiple LLIDTSP static simulations were generated (with known full-time-slot mission arrivals and topology changes).

[0091] For each example Extract the full-time-slot task requirements of example I. With network topology Then, the example I is decomposed into T time-slot subproblems. Each sub-problem corresponds to a time slot state in a dynamic scenario. .

[0092] For each time slot subproblem Extracting state features The solution is obtained using a two-stage temporal heuristic method, which extracts the link selection scheme. The true value of the corresponding link This forms a "status-label" pair. .

[0093] The result Add D and return dataset D. Summarize all "state-label" pairs and split them into a training set (80%) and a validation set (20%). This provides decision labels for each example in the dataset, forming a decision training set.

[0094] (2) Supervised learning In this embodiment, a supervised learning method is used to train the model. Based on the labels generated by the two-stage temporal heuristic, the machine learning model is trained by minimizing the difference between the predicted link value and the label value in each time slot. The true link value output by the two-stage temporal heuristic in a static example is shown. For the labels, minimize the mean squared error loss:

[0095] in, This represents the average number of links per time slot. This is the predicted link value. The training method is as follows: Input: Training dataset D, initial model M (including parameters) ), Training epochs E, Batch size B. Output: The trained model M.

[0096] First, the data is split: the training dataset is randomly divided into a training set (80%, used for parameter updates) and a validation set (20%, used for generalization ability evaluation) in a 4:1 ratio. Next, initialize the parameters: set the learning rate (to control the magnitude of parameter updates), the convergence threshold (to determine if the model has converged), and the iteration counter (initially set to 0 to record the number of training rounds). Iterative training loop: If the iteration counter is less than the preset number of training rounds and the change in validation loss is greater than the convergence threshold, repeat the following operation: Shuffle the training set. And divide the samples according to batch size Batch training: For each batch Extract sample status within the batch With tags →Model Prediction → Calculate the loss → Backpropagation updates model parameters Calculate the current validation loss using the validation set. The iteration counter is incremented by 1. For learning rate, Indicates parameters The change Let k represent the validation set, and k be the number of batches for the sample partitioning.

[0097] Output model: After the loop terminates, output the trained model M (the parameters are optimized and it has the ability to generalize prediction).

[0098] Next, we will use simulation experiments to compare and demonstrate the effectiveness of this solution with actual examples.

[0099] This embodiment constructs an emergency data transmission scenario using a low-Earth orbit constellation through simulation and designs comparative experiments to evaluate model performance. Since LLHVEDTP lacks publicly available mature datasets, this embodiment uses a static prospective problem data generation method to construct training and testing datasets. This embodiment adopts the Walker constellation configuration, containing 400 satellites at an orbital altitude of 1000 km and an orbital inclination of 97.86°. There are four ground stations located at... , A network topology diagram was generated using simulation software. The time slot length was set to 200 seconds, and the simulation planning time was 80 minutes (24 time slots). The relevant satellite parameters were set as follows: , Satellite downlink capacity is The value is randomly selected from Mb / s, while the inter-satellite link capacity is set to 10 Mb / s. The arrival of mission data uses vectors. The arrival times follow a normal distribution, where m represents the task number. Indicates the amount of data for the task, and follows the rules of data storage. The data is uniformly distributed, where s represents the satellite number from which mission data can be acquired, t represents the arrival time slot of the mission, and the value of each mission ranges from [0,1] and follows a normal distribution. The task data volume and the number of tasks in each time slot follow a Poisson distribution with a mean of 10. A total of 1000 static examples were generated, of which 800 were used as the training set and the other 200 as the test set. The average reward of the 200 static examples was used as the evaluation metric.

[0100] All simulation experiments were conducted under a unified hardware environment.

[0101] In this embodiment, the link value predictor uses a graph neural network as the link value prediction model with a two-stage optimization model framework.

[0102] To verify the effectiveness of this model, the following three commonly used algorithms in the field of dynamic scheduling are selected as the baseline for the link establishment strategy: Greedy algorithm: a heuristic algorithm that prioritizes the link with the largest capacity in each time slot and transmits according to task priority from high to low, simulating the "local optimum" decision commonly used in engineering. Random algorithm: Randomly selects links to transmit data, serving as a performance lower limit reference and verifying the effectiveness boundary of the algorithm; Rolling Horizon Algorithm: A phased optimization method based on dynamic programming, which divides every 3 time slots into 1 window (for optimal performance) and performs local optimization within each window; Table 2 shows the comparison results of the average total task reward of each algorithm on the test set and the difference between the results and the labels given by the Temporal Two-Stage Heuristic Method (STPHA). It can be seen that the average task reward of our proposed scheme (MLETOM) (20.01) is significantly higher than other dynamic algorithms, approaching 4 / 5 of the upper bound solution of STPHA. The differences between the random algorithm and the greedy algorithm and STPHA both exceed 45%, and the goodness of their solutions exceeds 25%. This indicates that our proposed scheme effectively approximates the static optimal solution through "prediction + optimization." The distribution of the total task reward is as follows: Figure 4 As shown, MLETOM's returns are concentrated and its upper quartile is close to STPHA, indicating that it performs stably in different cases and is more robust than other algorithms.

[0103] Table 2. Performance Comparison of Different Algorithms with STPHA

[0104] To further verify MLETOM's dynamic adaptability, we compared the gains of each algorithm in new tasks over 24 time slots (the total value of newly completed tasks in each time slot).

[0105] like Figure 5 As shown, the proposed scheme exhibits relatively stable incremental gains across all time slots, consistently outperforming both greedy and random algorithms. Particularly in time slots 10-20, the incremental gains advantage reaches 30%-50%, demonstrating its effective ability to handle sudden task arrivals. Compared to the rolling time-domain method's sharp increase in the last time slot, this scheme demonstrates greater robustness and balance. The results indicate that this scheme can effectively learn strategies from optimal solutions, ensuring complete transmission of the corresponding tasks in each newly arriving task set, thus providing a more stable and higher-quality solution.

[0106] The cumulative evolution of total task rewards over time is as follows: Figure 6 As shown, the cumulative total revenue curve of this scheme is consistently ahead, significantly higher than other comparative algorithms, and the gap with other algorithms continues to widen in the later stages (after time slot 15). This indicates that through the synergy of "link value prediction - data optimization allocation", it avoids the performance degradation caused by resource exhaustion or link selection errors in traditional algorithms. This fully demonstrates that this scheme can better adapt to the changing task requirements in dynamic environments, thereby achieving higher total task revenue.

[0107] In yet another embodiment, this solution can also be implemented systematically, the system comprising: The spatiotemporal evolution graph construction module is used to embed constellation resource status as node attributes and synchronously describe topological connectivity and resource availability in order to construct a spatiotemporal evolution graph model. The initial setup module is used to determine the decision variables and objective function of the task planning problem, and to define the constraints. The two-stage optimization model module constructs and stores the trained two-stage optimization model, and plans the low-Earth orbit constellation emergency data transmission mission based on the trained two-stage optimization model. The model training module is used to construct the training dataset and train the two-stage optimization model through supervised learning.

[0108] Furthermore, the system may also include a raw data input module, through which users can input the raw constellation dataset that needs to be predicted and used to obtain emergency data transmission mission planning schemes into the system.

[0109] Furthermore, the system may also include an output module to output the system's prediction results data for easy viewing or use by users. This output can be in various formats, such as dynamic image representation or dataset format.

[0110] Furthermore, the present invention can also be implemented as a device, i.e., configured as an electronic device. The device includes one or more processors and a memory.

[0111] The processor may be a central processing unit (CPU) or other processing unit with data processing and / or information execution capabilities, and may control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc. One or more computer program information may be stored on the computer-readable storage medium, and the processor may run the program information to implement the methods or other desired functions provided in the various embodiments of the present invention described above.

[0112] In one example, the device may also include input and output devices, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0113] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this solution includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which the embodiments of this solution pertain. The processor performs the various methods and processes described above. For example, the method embodiments of this solution can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for planning emergency data transmission missions for low-Earth orbit constellations, characterized in that: The method includes: S1. Embed constellation resource status as node attributes and describe topological connectivity and resource availability simultaneously to construct a spatiotemporal evolution graph model. S2. Determine the decision variables and objective function for the task planning problem, and define the constraints. S3. Construct a two-stage optimization model and a training dataset. Train the two-stage optimization model using supervised learning to obtain the trained two-stage optimization model. S4. Based on the trained two-stage optimization model, plan the emergency data transmission mission for the low-Earth orbit constellation.

2. The method according to claim 1, characterized in that, In S1, the specific method for constructing the spatiotemporal evolution diagram model is as follows: First, the overall planning duration is broken down into T consecutive time slots. Each time slot t corresponds to a static subgraph. , where the set of nodes Includes satellite nodes With ground station nodes N is the total number of satellites in the constellation, and M is the total number of ground stations, satisfying the following conditions: , Represents the set of potential links. Indicates resource attributes; Secondly, in time slots Inside, All remain unchanged, and this time slot is regarded as a static network slice, within which link selection and data allocation decisions are made; Furthermore, when transitioning from time slot t to t+1, the satellite's orbital motion causes... The relative position changes of satellite nodes in the middle, which in turn trigger The update, and the data transmission within time slot t, will consume satellite buffer resources, making... based on The amount of data transmitted is dynamically updated. Furthermore, potential links between satellites exist only within the visible window, and active links are filtered through a connection plan; the connection plan is defined as a set of conflict-free connections after scheduling, which is a subset of potential links.

3. The method according to claim 1, characterized in that, In S2, the decision variables include: Link connectivity variables represented by binary integer variables , indicating the connectivity status of link e within time slot t; Data transfer variables This represents the amount of data transmitted in task f within link e in time slot t. , The set of data transmission tasks arriving in time slot t; Task completion rate represented by a binary integer variable , indicating time slot The degree of completion of task f at the end.

4. The method according to claim 1, characterized in that, In S2, the optimization objective is to maximize the value of the complete mission data transmitted back to the ground station up to the current time slot t. Therefore, the specific objective function is: in, The value of task f; The completion rate of task f. This is the set of data transmission tasks arriving in time slot t.

5. The method according to claim 1, characterized in that, In S2, the constraints include: connection quantity constraints, data flow constraints, cache constraints, energy constraints, and task integrity constraints. The data flow constraints include link capacity limit constraints, task data arrival constraints, traffic conservation constraints, and ground station aggregation constraints. The energy constraints include solar charging constraints and energy balance constraints after determining the energy consumption for data forwarding, data receiving, and basic operation.

6. The method according to claim 1, characterized in that, In S3, the two-stage optimization model includes an input module, a chain building scheme generation module, and an optimization module; The input module is used to determine various input characteristics in the system state and to determine the decision evolution rules; The chain building scheme generation module is used to map the system state to the link value through supervised learning, thereby obtaining a set of feasible links; The optimization module optimizes the set of feasible links using linear programming to select the optimal link plan.

7. The method according to claim 6, characterized in that, The input features include: Satellite status characteristics N represents the number of satellites, and K represents the number of missions; mission status characteristics Network state characteristics M represents the number of ground stations, and the element This indicates that node i (satellite / ground station) and node j are visible in time slot t and can establish a link; otherwise, it is 0. The above features together constitute the state feature vector: .

8. The method according to claim 1, characterized in that, In S3, training the two-stage optimization model first requires constructing a training dataset, which is done as follows: First, initialize the training dataset. The simulation generates multiple static cases, forming a static case set I. All static cases have known full-time-slot task arrival and topology changes. Then, for each case in the case set, the full-time-slot task requirements for that case are extracted. With network topology The example is then decomposed into T time-slot subproblems. Each sub-problem corresponds to a time slot state in a dynamic scenario. ; Secondly, for each time slot subproblem Extracting state feature vectors The two-stage temporal heuristic method is used to solve the problem and extract the link selection scheme. The true value of the corresponding link Forming state-label pairs ; Finally, the result will be Add D and return dataset D; dataset D is used as the training dataset, divided into training set and validation set for model training.

9. A low-Earth orbit constellation emergency data transmission mission planning system, characterized in that: The system includes: The spatiotemporal evolution graph construction module is used to embed constellation resource status as node attributes and synchronously describe topological connectivity and resource availability in order to construct a spatiotemporal evolution graph model. The initial setup module is used to determine the decision variables and objective function of the task planning problem, and to define the constraints. The two-stage optimization model module constructs and stores the trained two-stage optimization model, and plans the low-Earth orbit constellation emergency data transmission mission based on the trained two-stage optimization model. The model training module is used to construct the training dataset and train the two-stage optimization model through supervised learning.

10. Low-Earth Orbit Constellation Emergency Data Transmission Mission Planning Equipment, characterized in that, The device includes at least a storage device and a processor, wherein the processor invokes data and computer instructions in the storage device to execute the low-Earth orbit constellation emergency data transmission mission planning method as described in any one of claims 1-8.

Images