Link state aware routing optimization method for dynamic low earth orbit satellite networks

Abstract

The application discloses a link state-aware routing optimization method for a dynamic low-orbit satellite network and belongs to the technical field of satellite communication networks.The method first constructs a dynamic topology model of a low-orbit satellite constellation;uniformly represents satellite topology connection relationships and link state parameters under different time slots and constructs a space-time state sample according to a historical time window;then generates a candidate next-hop node supervision signal, trains a space-time learning routing model aware of link states; and finally uses the trained space-time learning routing model for online hop-by-hop routing decision-making.The application fuses dynamic topology information and link state features, mines the historical evolution law of link states by constructing a space-time learning model, realizes candidate next-hop node prediction, and combines real-time link states to guide the cost to complete online hop-by-hop routing decision-making, thereby improving the accuracy and stability of routing selection in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the fields of satellite communication networks, integrated air-space-ground networks, and routing optimization technology, and particularly to a link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks. Background Technology

[0002] Low Earth Orbit (LEO) satellite networks offer advantages such as wide coverage, low transmission latency, and strong global access capabilities. However, the high-speed motion of satellites causes dynamic and time-varying network topology and link states. Routing paths are prone to failure due to link switching, node movement, or local congestion, leading to frequent route reconfiguration, increased end-to-end latency, decreased throughput, and reduced routing success rate. Therefore, achieving efficient and stable routing decisions in dynamic environments has become an important research topic in the field of LEO satellite communications.

[0003] Existing routing optimization methods for LEO satellite networks mainly include topology-aware routing methods, link-state maintenance-based routing methods, path optimization methods based on static link costs or shortest paths, and improved routing methods oriented towards load balancing, traffic engineering, energy constraints, and congestion control. These methods can improve path selection capabilities in dynamic networks to some extent, but they typically rely on current topology information, preset link cost models, or fixed optimization rules. They fail to adequately utilize the historical correlation characteristics formed by link states evolving over time. When topology changes rapidly, service load surges, or constellation size expands, problems such as delayed routing decisions, frequent path switching, and insufficient stability can easily arise.

[0004] In recent years, intelligent methods such as deep learning, reinforcement learning, graph neural networks, and recurrent neural networks have been increasingly applied to routing optimization in dynamic satellite networks. These methods can learn the mapping relationship between network state and routing decisions using a data-driven approach, providing new technical ideas for intelligent routing in complex dynamic scenarios. However, most existing intelligent routing methods still suffer from problems such as insufficient modeling of the spatiotemporal correlation features of link states, inadequate integration of model output and real-time link state constraints, and poor stability of online decision-making. They also struggle to simultaneously achieve optimal routing success rate, end-to-end latency, network throughput, and scalability.

[0005] In summary, existing routing optimization technologies for low-Earth orbit satellite networks still have limitations when dealing with complex scenarios such as highly dynamic topologies, rapid evolution of link states, and uneven spatiotemporal distribution of service loads. They are unable to simultaneously meet the requirements of dynamic low-Earth orbit satellite networks for high routing success rates, low end-to-end latency, high throughput, and stable routing decisions. Summary of the Invention

[0006] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a link-state-aware routing optimization method for dynamic low-Earth orbit (LEO) satellite networks. This method aims to fully utilize dynamic topology information and link state evolution characteristics within historical time windows, and combine them with real-time link state constraints to achieve hop-by-hop dynamic routing decisions. This will improve the routing success rate of dynamic LEO satellite networks, reduce end-to-end transmission latency, enhance network throughput performance, and strengthen the adaptability and robustness of the routing method under different constellation sizes and complex dynamic scenarios.

[0007] Technical solution: The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks of the present invention includes the following steps:

[0008] Step 1: Based on the orbital motion characteristics of low-Earth orbit satellite nodes, constellation configuration, and relative positional changes among satellites, establish a network topology model that changes over time to describe the connection relationships between satellite nodes and between satellites and ground nodes at different times.

[0009] Step 2: Considering the dynamic changes in the inter-satellite link status over time in low-Earth orbit satellite networks, a comprehensive link cost function for routing decisions is established by taking into account network status parameters.

[0010] Step 3: Based on the historical time window, obtain dynamic topology information and link status information at continuous moments to form a spatiotemporal status sample; select several neighboring nodes of the current node according to the comprehensive link cost to form a candidate next hop set, and use the second hop node of the optimal path as a supervision signal.

[0011] Step 4: Construct and train a link state-aware spatiotemporal learning routing model, which includes a spatial feature encoding module, a time series modeling module, and an output decision module.

[0012] Step 5: Input the topology and link status information acquired in real time from the dynamic low-Earth orbit satellite network into the trained spatiotemporal learning routing model to obtain the prediction results of candidate next-hop nodes; combine the real-time link status guidance cost to complete the hop-by-hop dynamic routing decision under the conditions of link availability, path acyclicity, and maximum hop count constraints.

[0013] Preferably, step 1 includes:

[0014] Using the Walker constellation as the networking basis for the LEO satellite network, assuming the constellation contains a total of A set of satellite nodes is defined as follows: Discretize continuous time into equally spaced time slot sequences Each time slot corresponds to a network snapshot; for any satellite In the time slot The geocentric coordinates are represented as a three-dimensional vector. ;

[0015] Based on satellite positions, obtain any two satellites and In the time slot geometric distance;

[0016] Let the maximum communication distance be The link availability is then defined as:

[0017] ;

[0018] in, Represents any two satellites and In the time slot geometric distance, For time slots adjacency matrix middle element, calculation Generation time slot adjacency matrix ,and It is a symmetric matrix, that is ;

[0019] Therefore, time slot The LEO satellite network topology model is abstracted as a time-varying undirected graph. ,in, This indicates additional constraint terms, which are introduced for modification when pointing constraints, occlusion, or maximum degree limits need to be considered. Indicates a feasible link.

[0020] Preferably, step 2 includes:

[0021] For any time slot With any link Considering the propagation delay determined by geometric distance, the link propagation delay is defined as: ,in The speed of light;

[0022] For link carrying capacity, define link capacity. The capacity is reduced by distance and an upper limit is set. Reference distance , represented as:

[0023] ;

[0024] Consider arbitrary time slots Next from source node to the destination node A feasible path Represented as:

[0025] ;

[0026] The number of hops on the path is:

[0027] ;

[0028] in, Indicates the number of hops in the path. Used to enumerate each link on the path ;

[0029] The end-to-end propagation delay of this path is defined as the sum of the link propagation delays, expressed as:

[0030] ;

[0031] The bottleneck capacity describes the path throughput capacity, expressed as:

[0032] ;

[0033] In each time slot For each request Route selection is accomplished by minimizing the combined link cost function, which includes delay and throughput penalties, denoted as:

[0034] ;

[0035] In the formula, This is a tradeoff factor between latency and throughput; To transform the preference for larger bottleneck capacity into a penalty mapping function that minimizes cost, This indicates a maximum hop count constraint; constraint Ensure that each hop on the path selects an available link in the current time slot; constraints Acyclic constraints are used to avoid invalid paths caused by loop forwarding; constraints This is a constraint on the maximum number of hops.

[0036] Preferably, step 3 includes:

[0037] Define metadata for each sample. ,in and These represent the source node and the destination node, respectively. This represents the time index used to construct the history window;

[0038] At the current forwarding node Neighbor set Construct a candidate next-hop set and constrain its size. ,Right now ;when At that time, this metadata The sample will be deemed invalid and removed.

[0039] make Indicates snapshot Downlink The propagation delay Indicates link capacity. If it is a capacity-stable term, then from node to his neighbor The local cost is defined as:

[0040] ;

[0041] To all neighbors According to local cost Sort in ascending order, take the first few. A candidate set ;

[0042] For candidate set Reorder the data so that the labels are always represented as candidate indexes;

[0043] To standardize the numerical scale of different snapshots, the normalization factor for latency and capacity in each snapshot is defined as follows:

[0044] ;

[0045] ;

[0046] And normalize the latency and capacity respectively;

[0047] For any candidate action Let the candidate node be At every historical moment The above constructs a six-dimensional feature, represented as:

[0048] ;

[0049] The first three items represent the reachability and link quality from the current node to the candidate node, while the last three items represent the reachability and link quality from the candidate node to the destination node.

[0050] Six-dimensional features It is a time slice of the sample state tensor, Stacking them over time yields the sample state tensor. Each candidate action corresponds to a temporal feature sequence, which serves as a spatiotemporal state sample.

[0051] For each sample In snapshot above As an edge weight, construct a weighted graph and solve from arrive shortest path If the shortest path exists and its node sequence is The optimal next hop is defined as:

[0052] ;

[0053] And the optimal next hop As a category label ,like If the shortest path does not exist, the sample is deemed invalid and discarded, and the node supervision signal is constructed from the classification label.

[0054] Preferably, the training process of the spatiotemporal learning routing model includes:

[0055] sample state tensor The input is fed into the spatiotemporal learning routing model, where the sample state tensor... Expanded in the time dimension, it becomes a length of... The time-series state sequence, denoted as Each frame has a dimension of 1. ;

[0056] First, spatial encoding is performed on each time slice to select candidate dimensions. The relative differences and reachability structure within the candidate set are extracted using the main axis, and a low-dimensional embedding representation is output. Subsequently, a temporal encoder is used to process the length of the set. Modeling the embedded sequences yields spatiotemporal feature vectors for next-hop discrimination. Finally, in the output layer, for The conditional probability distribution of the candidate next hops is given as follows:

[0057] ;

[0058] in, Indicates in time slot Next, select the first from the candidate set. Each candidate node is used as the predicted probability of the next hop; and These are the standard learnable parameters of the Softmax layer.

[0059] Preferably, the training objective of the spatiotemporal learning routing model is:

[0060] ;

[0061] in, For the set of model parameters, For the training set sample index set, For category weights, For indicator functions, The candidate set category label is mapped from the first hop of the optimal route in the next time step. The candidate set category label is defined by the candidate node. This indicates that the candidate next hop is given a conditional probability distribution.

[0062] Preferably, step 5 includes:

[0063] First, the candidate set is selected from the currently available neighbor nodes. Composed of candidate next-hop nodes Construct sample state vector The input is fed into the trained spatiotemporal learning model to obtain the probability output. ;

[0064] In the next time slot Animated GIF Above, using the link proxy cost as the edge weight, let the link state bootstrapping cost be the cost of forwarding from the current node through the candidate node and continuing to the destination node:

[0065] ;

[0066] in, In the animated image The cost of the shortest path obtained by weighting the edges. This represents the local cost from the current forwarding node to the candidate next hop;

[0067] On the candidate set Normalization is performed to obtain And define the fusion scoring function:

[0068] ;

[0069] in, This represents the link state guidance strength coefficient. It is a very small constant;

[0070] Therefore, the highest-scoring move in the candidate set is selected as the next jump action:

[0071] ;

[0072] Link availability constraints are applied only from candidates. The selection of nodes ensures that the acyclic constraint is maintained by keeping a set of visited nodes. And exclude during selection The candidate implementations use a counter to control the path length for the maximum hop count constraint. ,when If the request is immediately terminated and deemed a failure, or if the candidate set is empty, all candidates are infeasible, or the shortest path is unreachable, termination is also triggered to ensure that the algorithm outputs a feasible path that satisfies the constraints or an explicit failure flag.

[0073] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0074] This invention integrates dynamic topology information and link state characteristics. By constructing a spatiotemporal learning model to mine the historical evolution patterns of link states, it achieves candidate next-hop node prediction and combines real-time link state-guided costs to complete online hop-by-hop routing decisions, thereby improving the accuracy and stability of routing selection in dynamic environments. Simulation results, conducted in Walker constellation scenarios of different scales, demonstrate that this invention can improve routing success rate, reduce end-to-end transmission latency, enhance network throughput performance, and exhibit good stability and adaptability in network scenarios of varying scales. Attached Figure Description

[0075] Figure 1 This is a diagram showing the construction process of samples and supervisory signals, as well as the state tensor structure.

[0076] Figure 2 This is a schematic diagram of the spatiotemporal learning model structure guided by link state;

[0077] Figure 3 This is a flowchart of the link-state guided spatiotemporal network routing algorithm LS-STNet;

[0078] Figure 4 This is a comparison of success rates when N=36;

[0079] Figure 5 This is a comparison of success rates when N=100;

[0080] Figure 6 This is a time delay comparison when N=36;

[0081] Figure 7 This is a time delay comparison when N=100;

[0082] Figure 8 This is a comparison of throughput when N=36;

[0083] Figure 9 This is a comparison of throughput when N=100. Detailed Implementation

[0084] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0085] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0086] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0087] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0088] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0089] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0090] The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks described in this embodiment includes the following steps:

[0091] Step 1: Based on the orbital motion characteristics of low-Earth orbit satellite nodes, constellation configuration, and relative positional changes among satellites, establish a network topology model that changes over time to describe the connection relationships between satellite nodes and between satellites and ground nodes at different times.

[0092] Further, step 1 includes:

[0093] Using the Walker constellation as the networking foundation for the LEO satellite network, ephemeris data is used to characterize the evolution of satellite positions over time, and a time-varying topology model is established based on this. It is assumed that there are a total of [missing information - likely satellites] in the constellation. A set of satellite nodes is defined as follows: Discretize continuous time into equally spaced time slot sequences Each time slot corresponds to a network snapshot; for any satellite In the time slot The geocentric coordinates are represented as a three-dimensional vector. The three-dimensional vector is given directly from the ephemeris file, and its coordinate system can be geocentric coordinates;

[0094] Based on satellite positions, obtain any two satellites and In the time slot The geometric distance is expressed as:

[0095] ;

[0096] Considering that the establishment and maintenance of ISLs are usually affected by constraints such as maximum communication distance and visibility, this example uses a communication distance threshold to abstract link availability, assuming the maximum communication distance is... The link availability is then defined as:

[0097] ;

[0098] in, For time slots adjacency matrix middle element, calculation Generation time slot adjacency matrix ,and It is a symmetric matrix, that is ;

[0099] Therefore, time slot The LEO satellite network topology model is abstracted as a time-varying undirected graph. ,in, This represents an additional constraint term, introduced to modify the set when pointing constraints, occlusion, or maximum degree limits need to be considered. Specifically, it removes links that do not meet the constraints from the set. Indicates a feasible link.

[0100] Due to the high-speed motion of the satellite, three-dimensional vectors Follow Changes, thus leading to changes in the adjacency matrix With additional constraints It also evolves over time. This time-varying graphical model can uniformly describe the topology snapshots of the LEO satellite network in different time slots, providing a foundation for subsequent link performance modeling and routing performance evaluation.

[0101] The correction method is illustrated using the maximum link establishment degree limit as an example. (Satellite) The current number of established links is The maximum number of links that can be established is This introduces additional terms to the original link cost. If... Furthermore, if the link is not an existing link, then it will be forcibly set. This makes the link no longer belong to Therefore, it is excluded from the set of feasible paths; when near At that time, priority can be given to reserving links corresponding to satellites with higher terminal idle rates. This causes the overall link cost function to naturally favor these satellites during calculation. Similarly, antenna pointing constraints can reduce the bandwidth of links outside the scanning range. Set to 0 to from Remove from the middle, occlusion constraints can be used to remove links that are occluded. Set to 0 to The removal of the middle element and the unification of all three methods involve adjusting the values ​​of the adjacency matrix elements to correct the issue. This leads to an update of the topology model. The processing methods are consistent.

[0102] Step 2: Considering the dynamic changes in the inter-satellite link status over time in low-Earth orbit satellite networks, a comprehensive link cost function for routing decisions is established by taking into account network status parameters.

[0103] Because the link transmission performance information acquired by dynamic low-Earth orbit satellite networks includes network state parameters such as propagation delay, link capacity, and link availability, but these parameters continuously evolve with satellite motion and the network topology connections differ across time slots, they cannot be directly used for training spatiotemporal learning models. Therefore, it is necessary to uniformly encode and structure the dynamic topology information and link state information to provide a data foundation for subsequent spatiotemporal state sample construction and candidate next-hop supervision signal generation.

[0104] Further, step 2 includes:

[0105] For any time slot With any link Considering the propagation delay determined by geometric distance, the link propagation delay is defined as: ,in For the speed of light; The node pair is unavailable and does not participate in path construction and metric calculation;

[0106] For link carrying capacity, define link capacity. The capacity is reduced by distance and an upper limit is set. Reference distance , represented as:

[0107] ;

[0108] Consider arbitrary time slots Next from source node to the destination node A feasible path Represented as:

[0109] ;

[0110] The number of hops on the path is:

[0111] ;

[0112] in, Indicates the number of hops in the path. Used to enumerate each link on the path ;

[0113] The end-to-end propagation delay of this path is defined as the sum of the link propagation delays, expressed as:

[0114] ;

[0115] The bottleneck capacity describes the path throughput capacity, expressed as:

[0116] ;

[0117] For discrete time slot sets Given a set of communication requests in any time slot The goal of routing decisions is to determine the dynamic topology of the current time slot. Select a path from the source node. to the destination node Feasible path According to the defined and To achieve coordinated optimization of low latency and high throughput while ensuring reachability and engineering constraints, a parameter trade-off is adopted. The multi-objective routing problem is transformed into a solvable single-objective form. In each time slot... For each request Routing is accomplished by minimizing the combined cost function, which includes delay and throughput penalties, denoted as:

[0118] ;

[0119] In the formula, This is a tradeoff factor between latency and throughput; To transform the preference for larger bottleneck capacity into a penalty mapping function that minimizes cost, This indicates a maximum hop count constraint; constraint To ensure that each hop on the path selects an available link in the current time slot, since Constrained by the link availability determination in step 1, this constraint implicitly includes physical visibility conditions such as maximum communication distance; Acyclic constraints are used to avoid invalid paths caused by loop forwarding; constraints For the maximum hop count constraint, its upper bound is... It should be consistent with the engineering implementation and simulation settings.

[0120] In one example A monotonically decreasing form can be adopted, thus ensuring that the larger the bottleneck capacity, the smaller the corresponding penalty and the lower the overall cost. For example, the following can be considered:

[0121] ,in, This represents the upper limit of the link capacity. The larger the bottleneck capacity, the smaller the penalty and the lower the overall cost.

[0122] Considering that inter-satellite links (ISLs) change rapidly over time, this invention further constructs a link state-aware learning strategy for slot-level decision-making based on the above-mentioned routing optimization objectives and feasible domain constraints, so as to achieve end-to-end routing optimization in dynamic scenarios.

[0123] Step 3: Based on the historical time window, obtain dynamic topology information and link status information at continuous moments to form a spatiotemporal status sample; select several neighboring nodes of the current node according to the comprehensive link cost to form a candidate next-hop set, and use the second-hop node of the optimal path as a supervision signal.

[0124] To enable the model to perform hop-by-hop next-hop selection in the dynamic topology of the LEO satellite network, routing decision-making is abstracted as a supervised learning problem. Given a sequence of historical link states and the current routing task... Under these conditions, the next-hop node is predicted. Since the ISL changes during the evaluation process, spatiotemporal samples are established using discrete snapshots. For a number of satellites... The constellation, let the total number of snapshots be . For any time Based on the length of the historical window Construct state inputs and utilize the next snapshot. The link state generates a supervision label, thereby ensuring that the label reflects the optimal next-hop selection for the topology in the next time step.

[0125] Furthermore, step 3 includes:

[0126] Define metadata for each sample. ,in and These represent the source node and the destination node, respectively. This represents the time index used to construct the history window;

[0127] To limit the action space to a controllable size and closely resemble the actual hop-by-hop routing process, at the current forwarding node Neighbor set Construct a candidate next-hop set and constrain its size. ,Right now ;when At that time, this metadata Invalid samples will be judged as invalid and removed, thereby ensuring that the output dimension of the subsequent network is fixed and avoiding training noise caused by insufficient candidates;

[0128] make Indicates snapshot Downlink The propagation delay Indicates link capacity. If it is a capacity-stable term, then from node to his neighbor The local cost is defined as:

[0129] ;

[0130] To all neighbors According to local cost Sort in ascending order, take the first few. A candidate set ;

[0131] To avoid the network learning order biases irrelevant to the actual decision due to a fixed candidate ranking, the candidate set is reordered without changing its contents. Rearrange the labels so that they are always expressed as candidate indices rather than absolute node numbers, thereby improving the reliability of model generalization.

[0132] Regarding state input, only the link state sequence and metadata are saved. The morphology is reconstructed as needed during training and evaluation. For each sample, within a historical window... Link state features related to candidate actions are extracted; to unify the numerical scale of different snapshots, the normalization factor for latency and capacity in each snapshot is defined as follows:

[0133] ;

[0134] ;

[0135] And normalize the latency and capacity respectively;

[0136] For any candidate action Let the candidate node be At every historical moment The above constructs a six-dimensional feature, represented as:

[0137] ;

[0138] The first three terms represent the reachability and link quality from the current node to the candidate node, while the last three terms represent the reachability and link quality from the candidate node to the destination node. These terms are used to explicitly introduce the routing task. Directional information;

[0139] Six-dimensional features It is a time slice of the sample state tensor, Stacking them over time yields the sample state tensor. Each candidate action corresponds to a temporal feature sequence, serving as a spatiotemporal state sample. This provides the subsequent spatiotemporal learning model with both candidate differences and historical evolution. The construction process of the samples and supervision signals, as well as the organization of the state tensor, are as follows: Figure 1 As shown.

[0140] For each sample In snapshot above As an edge weight, construct a weighted graph and solve from arrive shortest path If the shortest path exists and its node sequence is The optimal next hop is defined as:

[0141] ;

[0142] And the optimal next hop As a category label ,like If the shortest path does not exist, the sample is deemed invalid and discarded, and the node supervision signal is constructed from the classification label.

[0143] In this example, the training and test sets are divided by time, i.e., the previous time. The sample state tensors of one time index are used as the training set, and the rest are used as the test set. This avoids the simultaneous appearance of highly correlated samples from the same time slice in both training and testing, which could lead to data leakage and thus more objectively evaluate the model's ability to generalize to future topologies.

[0144] Step 4: Construct and train a link-state-aware spatiotemporal learning routing model, such as... Figure 2 As shown, the spatiotemporal learning routing model includes a spatial feature encoding module, a time series modeling module, and an output decision module.

[0145] Based on the previously constructed sample state tensor, an end-to-end spatiotemporal learning model, denoted as the LS-STNet model, is designed to map the link state evolution on the candidate next-hop set to the next-hop selection distribution, thereby providing learnable decision priors for online hop-by-hop routing.

[0146] Furthermore, the training process of the spatiotemporal learning routing model includes:

[0147] sample state tensor The input is fed into the spatiotemporal learning routing model, where the sample state tensor... Expanded in the time dimension, it becomes a length of... The time-series state sequence, denoted as Each frame has a dimension of 1. ;

[0148] First, spatial encoding is performed on each time slice to select candidate dimensions. The relative differences and reachability structure within the candidate set are extracted using the main axis, and a low-dimensional embedding representation is output. Subsequently, a temporal encoder is used to process the length of the set. The embedded sequences are modeled using techniques such as LSTM to capture the correlation and trend of link quality over time, resulting in a spatiotemporal feature vector for next-hop discrimination. Finally, in the output layer, for The conditional probability distribution of the candidate next hops is given as follows:

[0149] ;

[0150] in, Indicates in time slot Next, select the first from the candidate set. Each candidate node is used as the predicted probability of the next hop; and These are the standard learnable parameters of the Softmax layer.

[0151] Output of the spatiotemporal learning routing model It is not directly equivalent to the final routing decision, but rather serves as a learning prior combined with link state calculation. The former reflects the learnable spatiotemporal patterns in historical evolution, while the latter embodies the interpretable link cost under the current snapshot, thus forming an online hop-by-hop decision strategy that satisfies the constraints in the next section.

[0152] During the supervised learning phase, the labels are given by the supervision signal generation rules, which map the first hop of the optimal route in the next time step to the category label in the candidate set. .

[0153] Furthermore, the training objective of the spatiotemporal learning routing model is:

[0154] ;

[0155] in, For the set of model parameters, For the training set sample index set, For category weights, For indicator functions, The candidate set category label is mapped from the first hop of the optimal route in the next time step. The candidate set category label is defined by the candidate node. This indicates that the candidate next hop is given a conditional probability distribution.

[0156] Step 5: Input the topology and link status information acquired in real time from the dynamic low-Earth orbit satellite network into the trained spatiotemporal learning routing model to obtain the prediction results of candidate next-hop nodes; combine the real-time link status guidance cost to complete the hop-by-hop dynamic routing decision under the conditions of link availability, path acyclicity, and maximum hop count constraints.

[0157] Furthermore, step 5 includes:

[0158] First, the candidate set is selected from the currently available neighbor nodes. Composed of candidate next-hop nodes Construct sample state vector The input is fed into the trained spatiotemporal learning model to obtain the probability output. ;

[0159] In the next time slot Animated GIF Above, using the link proxy cost as the edge weight, let the link state bootstrapping cost be the cost of forwarding from the current node through the candidate node and continuing to the destination node:

[0160] ;

[0161] in, In the animated image The cost of the shortest path obtained by calculating the edge weights above. This represents the local cost from the current forwarding node to the candidate next hop;

[0162] On the candidate set Normalization is performed to obtain And define the fusion scoring function:

[0163] ;

[0164] in, This represents the link state guidance strength coefficient. It is a very small constant;

[0165] Therefore, the highest-scoring move in the candidate set is selected as the next jump action:

[0166] ;

[0167] They tend to reuse the stable selection preferences implied by historical spatiotemporal patterns, while This explicitly penalizes high-cost links and unreachable routes in the current snapshot, thus making the routing decision more aligned with the preset routing optimization objectives. To strictly satisfy the aforementioned constraints, online decision-making performs feasibility screening at each hop; link availability constraints are implemented by selecting candidates only from... The selection of nodes ensures that the acyclic constraint is maintained by keeping a set of visited nodes. And exclude during selection The candidate implementations use a counter to control the path length for the maximum hop count constraint. ,when The request is immediately terminated and deemed a failure. Termination is also triggered if the candidate set is empty, all candidates are infeasible, or the shortest path is unreachable. This ensures the algorithm outputs a feasible path that satisfies the constraints or a clear failure flag, maintaining consistency with subsequent evaluation metrics such as success rate, end-to-end latency, and throughput. To facilitate explanation of the online hop-by-hop routing decision process of this invention, the process can be represented as an algorithm flowchart. Figure 3 The process takes the source node, destination node, current network topology, link state information, and a trained LS-STNet model as input. Through candidate next-hop node prediction, link state-guided cost calculation, constraint screening, and path hop-by-hop update, a dynamic routing path from the source node to the destination node is obtained.

[0168] This invention uses the Walker constellation as the networking foundation for an LEO satellite network, utilizing ephemeris data to characterize the evolution of satellite node positions over time, and constructing a dynamic topology model and a link state model based on this. The constructed data samples include dynamic topology information, link state information, routing request information, historical window length, candidate next-hop set size, maximum number of hops, fusion coefficients of model output and link state cost, and minimal constants used for numerical stabilization. First, a unified representation of satellite topology connections and link state parameters under different time slots is required, and spatiotemporal state samples are constructed according to historical time windows. Then, candidate next-hop node supervision signals are generated to train a link state-aware spatiotemporal learning routing model. Finally, the trained LS-STNet model is used for online hop-by-hop routing decisions.

[0169] To verify the effectiveness of the proposed method, Walker constellation simulation scenarios of different scales were constructed, and the LS-STNet model proposed in this invention was compared with LSTM (Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735-1780.), CNN-LSTM (Shi, X.,Chen, Z., Wang, H., Yeung, DY, Wong, WK, & Woo, WC (2015). Convolutional LSTM Network: A Machine Learning Approach for PrecipitationNowcasting. Advances in Neural Information Processing Systems (NeurIPS), 28,802-810.) and DRL (Mao, B., Fadlullah, ZM, Tang, F., Kato, N., et al. (2018). Routing or Computing? The Paradigm Shift Towards Intelligent Computer Network Packet Transmission Based on Deep Learning. IEEE Transactions on Computers, 66(11), 1946-1960.) and other typical intelligent routing algorithms, the routing performance of the algorithms is evaluated by selecting indicators such as routing success rate, end-to-end transmission delay and network throughput. LSTM captures the long-term dependencies of temporal data through a gating mechanism; CNN-LSTM uses CNN to extract spatial features and LSTM to model temporal evolution, forming a spatiotemporal joint feature extraction structure; DRL learns adaptive routing strategies through agent-environment interaction, modeling routing decisions as a state-action-reward Markov decision process.

[0170] In the simulation evaluation, the performance of online hop-by-hop routing is quantified from three dimensions: reachability, latency, and throughput. For any time slot... For routing requests, the algorithm outputs a hop-by-hop forwarding sequence. If the sequence eventually arrives and satisfies all constraints, the request is considered successful; otherwise, it is considered a failure. The success rate is defined as the percentage of successful requests in the test set. For successful samples, end-to-end latency is represented by path propagation latency, and the average end-to-end latency is taken as the average end-to-end latency metric across the test samples. Throughput is represented by path bottleneck capacity, and the average throughput metric is ultimately taken as the average of all test samples.

[0171] This example further statistically analyzes the average comprehensive cost function values ​​of different algorithms under two constellation scales, and the results are shown in Table 1. This index is derived from the formula... The mean value of the defined comprehensive cost on successful samples is obtained. The smaller the value, the better the algorithm's overall performance in terms of latency and throughput.

[0172] Table 1. Comparison of average integrated objective function values ​​of different algorithms under different constellation sizes.

[0173]

[0174] As can be seen from Table 1, in and Under both constellation scales, LS-STNet achieves the minimum average integrated objective function value, demonstrating that this invention achieves superior overall routing performance under a unified multi-objective trade-off criterion. In this scenario, LS-STNet reduces the average integrated objective function value by approximately 45.7%, 48.4%, and 20.1% compared to CNN-LSTM, LSTM, and DRL, respectively; In the scenario, the latency was reduced by approximately 10.2%, 14.5%, and 34.3%, respectively. This indicates that the combination of the link state guidance mechanism and the spatiotemporal learning model can not only reduce end-to-end propagation latency, but also improve path throughput by suppressing bottleneck link risks, thus exhibiting better steady-state performance in terms of overall cost.

[0175] Figure 4 and Figure 5 The constellation sizes are given as follows: and The simulation shows how the success rate of different algorithms on end-to-end routing tasks changes over time. Overall, In this scenario, the success rates of all methods were high, but varying degrees of stability differences were still observed. The LS-STNet curve quickly entered a stable region after the initial stage with relatively small fluctuations, indicating that its hop-by-hop decision-making is more adaptable to time-varying topologies. CNN-LSTM and LSTM showed more pronounced fluctuations in certain time periods, suggesting that relying solely on temporal learning or shallow spatial representations is insufficient in characterizing changes in link availability and cost. The DRL curve was generally flat but exhibited periodic declines, reflecting that online selection based on value estimation may be affected by estimation errors and policy instability in dynamic topologies. When the constellation size increases to... At that time, the differences between methods widened significantly. LS-STNet maintained a high and relatively stable success rate, while the success rates of CNN-LSTM, LSTM, and DRL declined more significantly. This is because the candidate next-hop space increases, the number of path hops and the length of the decision chain increase after scaling, and the single-step decision error is more likely to accumulate and amplify in the end-to-end process. In contrast, LS-STNet further introduces link state cost constraints on the learned spatiotemporal decision priors, which helps to suppress infeasible or high-cost forwarding choices in the online inference stage, thus showing better reachability guarantees and scale robustness in large-scale dynamic networks.

[0176] Figure 6 For different algorithms in The latency comparisons showed fluctuations in all cases during the initial simulation phase, primarily due to frequent route adjustments caused by the rapid evolution of the constellation topology in the initial stage. As time progressed, the curves gradually converged and entered a steady-state region. In the steady-state phase, the average end-to-end latency of LS-STNet remained around 40ms, significantly lower than CNN-LSTM, LSTM, and DRL, with corresponding relative latency reductions of approximately 25.9%, 27.3%, and 21.6%, respectively. This indicates that in small-to-medium-scale constellations, introducing link-state costs and combining them with spatiotemporal feature learning can more stably select low-cost hop-by-hop forwarding directions, thereby reducing detours and unnecessary path adjustments caused by local misjudgments.

[0177] Figure 7The study describes how, with increasing network size, the average path length and the number of available candidate neighbors lead to an overall increase in steady-state latency for all methods, with more pronounced differences between methods. Within the steady-state range, LS-STNet maintains the lowest average latency, achieving relative reductions of approximately 11.6% and 14.7% compared to CNN-LSTM and LSTM, respectively; and a reduction of approximately 30.3% compared to DRL. The results show that pure learning methods are more prone to decision bias in large-scale action spaces and more complex topology evolutions, while the link-state guidance mechanism, by explicitly characterizing the combined costs of latency and capacity, provides consistent physical constraints and correction directions for the model output, enabling it to maintain lower latency and stronger steady-state robustness even with increased scale.

[0178] from Figure 8 It can be seen from this that, In this scenario, the throughput of all four methods stabilized after initial brief fluctuations. It can be observed that the normalized bottleneck capacity of the LS-STNet proposed in this invention remained stable at approximately 0.043~0.045 over the long term, significantly higher than the comparative methods. This represents an improvement of approximately 20%~25% compared to DRL; approximately 170%~190% compared to CNN-LSTM; and approximately 180%~200% compared to LSTM. This indicates that in small-scale constellations, relying solely on temporal prediction or value function approximation is more likely to favor locally low-cost links during hop-by-hop selection, thus creating smaller bottleneck capacities on the path. In contrast, the link state guidance mechanism can explicitly suppress bottleneck risks between candidate next hops, continuously improving the weakest link capacity of the end-to-end path, resulting in a higher and more stable throughput curve.

[0179] from Figure 9 It can be seen that, In scenarios with higher network redundancy and richer candidate paths, the overall performance of each method further improves in smaller-scale scenarios and enters a stable range within a short time. The LS-STNet of this invention gradually stabilizes at 0.092~0.094 after about 100 seconds, maintaining the highest throughput level overall; compared to the stable stage of CNN-LSTM (approximately 0.088~0.091), it still has a gain of about 3%~6%; compared to LSTM, it improves by about 15%~20%; and compared to DRL, it improves by about 80%~95%. Simultaneously, it can be observed that CNN-LSTM performs significantly better than LS-STNet at this scale, indicating that even relying solely on the learning model in large-scale constellations can obtain feasible hop-by-hop decisions from richer connectivity relationships; however, LS-STNet still re-evaluates candidate actions by fusing link state costs, more effectively avoiding end-to-end bottleneck links in the long term, maintaining its throughput advantage with less fluctuation and more stable performance.

[0180] In summary, this invention, starting from the routing optimization requirements of dynamic low-Earth orbit (LEO) satellite networks, analyzes the shortcomings of existing traditional and intelligent routing methods for LEO satellite networks in terms of dynamic topology adaptation, utilization of spatiotemporal correlation features of link states, and stability of online decision-making. Based on this, it proposes a link-state-aware spatiotemporal learning routing optimization method for dynamic LEO satellite networks. This method integrates dynamic topology information and link state features to construct spatiotemporal state samples and candidate next-hop supervision signals, and combines the LS-STNet model and real-time link state guidance costs to achieve online hop-by-hop routing decisions. This method, to a certain extent, solves the problems of insufficient utilization of historical evolution patterns of link states and poor routing stability in dynamic environments in existing routing methods, improving routing success rate, reducing end-to-end transmission latency, and enhancing network throughput performance. Therefore, the link-state-aware spatiotemporal learning routing optimization method proposed in this invention is feasible and effective.

Claims

1. A link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks, characterized in that, Includes the following steps: Step 1: Based on the orbital motion characteristics of low-Earth orbit satellite nodes, constellation configuration, and relative positional changes among satellites, establish a network topology model that changes over time to describe the connection relationships between satellite nodes and between satellites and ground nodes at different times. Step 2: Considering the dynamic changes in the inter-satellite link status over time in low-Earth orbit satellite networks, a comprehensive link cost function for routing decisions is established by taking into account network status parameters. Step 3: Based on the historical time window, obtain dynamic topology information and link status information at continuous moments to form a spatiotemporal status sample; select several neighboring nodes of the current node according to the comprehensive link cost to form a candidate next hop set, and use the second hop node of the optimal path as a supervision signal. Step 4: Construct and train a link state-aware spatiotemporal learning routing model, which includes a spatial feature encoding module, a time series modeling module, and an output decision module. Step 5: Input the topology and link status information acquired in real time from the dynamic low-Earth orbit satellite network into the trained spatiotemporal learning routing model to obtain the prediction results of candidate next-hop nodes; combine the real-time link status guidance cost to complete the hop-by-hop dynamic routing decision under the conditions of link availability, path acyclicity, and maximum hop count constraints.

2. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 1, characterized in that, Step 1 includes: Using the Walker constellation as the networking basis for the LEO satellite network, assuming the constellation contains a total of A set of satellite nodes is defined as follows: Discretize continuous time into equally spaced time slot sequences Each time slot corresponds to a network snapshot; for any satellite In the time slot The geocentric coordinates are represented as a three-dimensional vector. ; Based on satellite positions, obtain any two satellites and In the time slot geometric distance; Let the maximum communication distance be The link availability is then defined as: ； in, Represents any two satellites and In the time slot geometric distance, For time slots adjacency matrix middle element, calculation Generation time slot adjacency matrix ,and It is a symmetric matrix, that is ; Therefore, time slot The LEO satellite network topology model is abstracted as a time-varying undirected graph. ,in, This indicates additional constraint terms, which are introduced for modification when pointing constraints, occlusion, or maximum degree limits need to be considered. Indicates a feasible link.

3. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 2, characterized in that, Step 2 includes: For any time slot With any link Considering the propagation delay determined by geometric distance, the link propagation delay is defined as: ,in The speed of light; For link carrying capacity, define link capacity. The capacity is reduced by distance and an upper limit is set. Reference distance , is represented as: ； Consider arbitrary time slots Next from source node to the destination node A feasible path Represented as: ； The number of hops on the path is: ； in, Indicates the number of hops in the path. Used to enumerate each link on the path ; The end-to-end propagation delay of this path is defined as the sum of the link propagation delays, expressed as: ； The bottleneck capacity describes the path throughput capacity, expressed as: ； In each time slot For each request Route selection is accomplished by minimizing the combined link cost function, which includes delay and throughput penalties, denoted as: ； In the formula, This is a tradeoff factor between latency and throughput; To transform the preference for larger bottleneck capacity into a penalty mapping function that minimizes cost, This indicates a maximum hop count constraint; constraint Ensure that each hop on the path selects an available link in the current time slot; constraints Acyclic constraints are used to avoid invalid paths caused by loop forwarding; constraints This is a constraint on the maximum number of hops.

4. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 3, characterized in that, Step 3 includes: Define metadata for each sample. ,in and These represent the source node and the destination node, respectively. This represents the time index used to construct the history window; At the current forwarding node Neighbor set Construct a candidate next-hop set and constrain its size. ,Right now ;when At that time, this metadata The sample will be deemed invalid and removed. make Indicates snapshot Downlink The propagation delay Indicates link capacity. If it is a capacity-stable term, then from node to his neighbor The local cost is defined as: ； To all neighbors According to local cost Sort in ascending order, take the first few. A candidate set ; For candidate set Reorder the tags so that they are always represented as candidate indexes; To standardize the numerical scale of different snapshots, the normalization factor for latency and capacity in each snapshot is defined as follows: ； ； And normalize the latency and capacity respectively; For any candidate action Let the candidate node be At every historical moment The above constructs a six-dimensional feature, represented as: ； The first three items represent the reachability and link quality from the current node to the candidate node, while the last three items represent the reachability and link quality from the candidate node to the destination node. Six-dimensional features It is a time slice of the sample state tensor, Stacking them over time yields the sample state tensor. Each candidate action corresponds to a temporal feature sequence, which serves as a spatiotemporal state sample. For each sample In snapshot above As an edge weight, construct a weighted graph and solve from arrive shortest path If the shortest path exists and its node sequence is The optimal next hop is defined as: ； And the optimal next hop As a category label ,like If the shortest path does not exist, the sample is deemed invalid and discarded, and the node supervision signal is constructed from the classification label.

5. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 4, characterized in that, The training process of the spatiotemporal learning routing model includes: sample state tensor The input is fed into the spatiotemporal learning routing model, where the sample state tensor... Expanded in the time dimension, it becomes a length of... The time-series state sequence, denoted as Each frame has a dimension of 1. ; First, spatial encoding is performed on each time slice to select candidate dimensions. The relative differences and reachability structure within the candidate set are extracted using the main axis, and a low-dimensional embedding representation is output. Subsequently, a temporal encoder is used to process the length of the set. Modeling the embedded sequences yields spatiotemporal feature vectors for next-hop discrimination. Finally, in the output layer, for The conditional probability distribution of the candidate next hops is given as follows: ； in, Indicates in time slot Next, select the first from the candidate set. Each candidate node is used as the predicted probability of the next hop; and These are the standard learnable parameters of the Softmax layer.

6. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 1 or 5, characterized in that, The training objective of the spatiotemporal learning routing model is: ； in, For the set of model parameters, For the training set sample index set, For category weights, For indicator functions, The candidate set category label is mapped from the first hop of the optimal route in the next time step. The candidate set category label is defined by the candidate node. This indicates that the candidate next hop is given a conditional probability distribution.

7. The link-state-aware routing optimization method for dynamic low-Earth orbit satellite networks according to claim 6, characterized in that, Step 5 includes: First, the candidate set is selected from the currently available neighbor nodes. Composed of candidate next-hop nodes Construct sample state vector The input is fed into the trained spatiotemporal learning model to obtain the probability output. ; In the next time slot Animated GIF Above, using the link proxy cost as the edge weight, let the link state bootstrapping cost be the cost of forwarding from the current node through the candidate node and continuing to the destination node: ； in, In the animated image The cost of the shortest path obtained by calculating the edge weights above. This represents the local cost from the current forwarding node to the candidate next hop; On the candidate set Normalization is performed to obtain And define the fusion scoring function: ； in, This represents the link state guidance strength coefficient. It is a very small constant; Therefore, the highest-scoring move in the candidate set is selected as the next jump action: ； Link availability constraints are applied only from candidates. The selection of nodes ensures that the acyclic constraint is maintained by keeping a set of visited nodes. And exclude during selection The candidate implementations use a counter to control the path length for the maximum hop count constraint. ,when If the request is immediately terminated and deemed a failure, or if the candidate set is empty, all candidates are infeasible, or the shortest path is unreachable, termination is also triggered to ensure that the algorithm outputs a feasible path that satisfies the constraints or an explicit failure flag.

Images