Low-altitude deterministic routing method based on in-band telemetering and high-order graph learning

Abstract

The invention provides a low-altitude deterministic routing method based on in-band telemetering and high-order map learning, and relates to the technical field of low-altitude internet of intelligence, the method comprises the following steps: in a data acquisition layer, detecting a full-link fine-grained path of a low-altitude network by using an auxiliary probe mechanism oriented to global link coverage optimization to obtain INT data; in the network modeling layer, on the basis of INT data, the probability of node and link faults is calculated by using a network fault prediction mechanism based on a high-order time sequence diagram neural network; and in a routing decision-making layer, based on the fault probability obtained through calculation, an optimal deterministic path is obtained by using a low-altitude deterministic routing mechanism based on fault probability minimization. According to the invention, the problems of communication quality reduction and interruption caused by rapid change of network topology in the low-altitude communication network are solved, and the reliability and transmission efficiency of the low-altitude network are improved.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude intelligent network technology, and particularly relates to a low-altitude deterministic routing method based on in-band telemetry and high-order graph learning. Background Technology

[0002] With the opening of low-altitude airspace, low-altitude communication networks have become an important component of future communication technology development. Low-altitude communication networks refer to a network system that utilizes aerial platforms such as drones, airborne base stations, and satellites to provide ground users with communication, positioning, monitoring, and data transmission services through wireless communication technology. This network system features high mobility, flexible deployment, and wide coverage, overcoming the limitations of traditional terrestrial communication networks, especially in complex environments, providing efficient wireless communication services. Low-altitude communication networks have broad application prospects in fields such as emergency rescue, intelligent transportation, environmental monitoring, and agricultural management.

[0003] However, low-altitude communication networks face many technical challenges in dynamic environments. First, low-altitude communication networks are highly dynamic and time-varying, with frequent changes in network topology and unstable communication links between nodes. This makes traditional communication protocols based on fixed topologies difficult to handle the complex low-altitude environment. Second, nodes in low-altitude networks are typically mobile platforms such as drones, whose high-speed movement introduces latency fluctuations and signal obstruction problems, placing higher demands on network reliability and real-time performance. Therefore, low-altitude communication networks must possess strong adaptability and robustness to cope with environmental changes and frequent fluctuations in network status.

[0004] In low-altitude communication networks, deterministic routing mechanisms can effectively handle frequent topology changes between nodes, optimize resource allocation, and ensure stable data transmission in complex and highly dynamic environments. Furthermore, network state awareness and network topology prediction can accurately perceive changes in network performance and provide prior knowledge of link availability, facilitating the early identification of potential network faults and bottlenecks, optimizing routing decisions, and reducing instability caused by link switching. Therefore, how to achieve accurate measurement and awareness of network state, accurate prediction of network topology, and reliable routing decisions in complex and dynamic low-altitude scenarios to ensure reliable transmission of low-altitude services is an important research topic.

[0005] To date, research on dynamic routing and network state prediction in low-altitude networks has mainly focused on two categories: one is graph neural network-based methods, used to capture the relationship between nodes and links over time and predict link states or network topology changes; the other is optimization scheduling methods based on high-precision network measurements, which use real-time network metrics for resource allocation and routing decisions. While graph neural network-based schemes can capture local node interactions and short-term dependencies, in low-altitude network scenarios with high-frequency topology changes, node memory modules need frequent updates, leading to a significant increase in memory and computational overhead, and their ability to model long-term dependencies is insufficient. Optimization scheduling schemes based on in-band telemetry can reflect network states in real time, but they typically only consider link states and resource allocation for single-hop nodes, lacking modeling of complex interactions between node groups, failing to effectively capture potential fault propagation patterns in the network, and being difficult to directly apply to deterministic routing strategies.

[0006] In terms of network state awareness, in-band network telemetry (INT), as a fine-grained technology for monitoring network performance, offers a new direction for network measurement in low-altitude scenarios due to its real-time performance and accuracy. The INT controller embeds telemetry commands into the service packets at the source node. During service packet transmission, the switch collects the corresponding network performance telemetry items according to the telemetry commands. Finally, at the destination node, the INT header and INT source data are extracted and submitted to the INT controller to generate a network telemetry report. INT achieves network performance data acquisition during the packet transmission process, allowing network administrators to directly capture network faults such as congestion control and configuration errors from the data plane.

[0007] Nodes and links in low-altitude communication networks often undergo frequent topology changes due to mobility and environmental variations. This makes traditional routing algorithms based on static topology inadequate for addressing network stability issues in highly dynamic environments. Topology prediction can effectively predict network change trends, providing a reliable basis for routing decisions and ensuring flexible and efficient real-time route adjustments in dynamic environments.

[0008] Currently, research on deterministic routing mechanisms has made some progress. Similar to low-altitude communication networks, satellite internet also needs to ensure communication stability and reliability in highly dynamic environments, while coping with frequent connection changes between nodes and signal obstruction. Therefore, research on deterministic routing in satellite networks can provide solutions for low-altitude communication networks. Summary of the Invention

[0009] To address the aforementioned shortcomings in existing technologies, this invention provides a low-altitude deterministic routing method based on in-band telemetry and high-order graph learning. This invention solves the problem of communication quality degradation and interruption caused by rapid changes in network topology in low-altitude communication networks, thereby improving the reliability and transmission efficiency of low-altitude networks.

[0010] To achieve the above objectives, the technical solution adopted by this invention is as follows: a low-altitude deterministic routing method based on in-band telemetry and high-order graph learning, comprising the following steps: S1. At the data acquisition layer, an auxiliary probe mechanism optimized for global link coverage is used to detect the fine-grained path of the entire link of the low-altitude network and obtain INT data. S2. In the network modeling layer, based on INT data, the probability of node and link failures is calculated using a network fault prediction mechanism based on a high-order temporal graph neural network. S3. At the routing decision layer, based on the calculated failure probability, the optimal deterministic path is obtained using a low-altitude deterministic routing mechanism that minimizes the failure probability.

[0011] Further, S1 includes the following steps: S101. Construct a bipartite graph of the low-altitude network topology based on candidate probe nodes and links to be observed, and use Hopcroft-Karp to obtain the maximum matching set of the bipartite graph. S102. Based on the maximum matching set, calculate the minimum edge-covered edge set; S103. By using the minimum edge to cover the edge set, INT data is obtained by generating probe message paths. In this process, a method of multiple probe nodes covering different edges simultaneously is adopted to improve the detection speed.

[0012] Furthermore, the method of using Hopcroft-Karp to find the maximum matching set of the bipartite graph is as follows: Based on the bipartite graph, a layered graph is constructed using breadth-first search and distances are calculated to find the length of the shortest augmenting path. Based on the length of the shortest augmenting path, a depth-first search is used to find augmenting paths where vertices do not intersect. Update the matching and loop until no augmenting path can be found in the breadth-first search, thus obtaining the maximum matching set of the bipartite graph.

[0013] Furthermore, S2 includes the following steps: S201. Preprocess the INT-based data; S202. Based on the preprocessed INT data, the topology of the low-altitude network at different timestamps is represented as a graph sequence; S203. Based on the graph sequence, perform high-order neighbor modeling on the nodes to complete the embedding update of the nodes; S204. Update the link features based on the embedding update results of the two end nodes; S205. Based on the update of link features and node features, the nodes and links of different time slices are stacked to form a time series feature sequence, and then input into a high-order time series graph neural network. S206. Use a high-order temporal graph neural network to predict the future dynamic characteristics of nodes or links; S207. Based on the predicted dynamic characteristics, predict the state at the next moment; S208. Input the predicted state of the next moment into the logistic regression model to calculate the probability of node and link failure.

[0014] Furthermore, the expressions for the probabilities of node and link failures are as follows:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] in, This represents the probability of a link failure. This represents the probability of node failure. The weights represent the link failure predictions. Indicates the link prediction status. Indicates the predicted state of the node. The weights represent the node failure predictions. and These represent the bias terms for link failure prediction and node failure prediction, respectively. For activation function, The mapping weights represent the link characteristics. Represents the dynamic characteristics of the connection. This indicates the bias in the link feature prediction. Represents the node feature mapping weights. Represents the dynamic characteristics of a node. This indicates the bias in node feature prediction. Indicates that the node is Dynamic characteristics at any given moment The time series representing the higher-order embeddings of nodes. Indicates the link is in Dynamic characteristics at any given moment This represents a time series composed of higher-order embeddings of edges. Represents a node exist Moment embedding, Represents a node exist Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in the th Layer embedding vectors, This represents a multilayer perceptron. Indicates the connection node and nodes The link in the th Layer embedding vectors, Indicates the first Layer nodes Embedded vector, Indicates the first Layer Neighbor weight matrix of order, Indicates the first Layer nodes Embedded vector, No. Layer nodes Embedded vector, Indicates the first The weight of the layer node itself, where K represents the maximum order of its higher-order neighbors. Indicates a higher-order neighbor. Indicates the first 1 node Represents a set of nodes.

[0026] Furthermore, step S3 includes the following steps: S301. The probability of node and link failure is mapped using a negative logarithm and quantified into an additive path cost. The path cost is accumulated segment by segment to obtain the cost map of the low-altitude network. S302. Based on the cost graph, generate multiple candidate paths; S303. Score and filter the candidate paths, and select the paths that meet the time delay constraints, jitter constraints and reliability constraints as feasible paths; S304. Define the comprehensive scoring function; S305. Use the comprehensive scoring function to obtain the optimal deterministic path among the feasible paths.

[0027] Furthermore, step S302 includes the following steps: Find the first shortest path and the shortest path Add to solution set middle; For the found Path Choose a fixed prefix, take the last node of the fixed prefix as the new starting point, find a new path again on the temporary graph after removing some edges or nodes, and put it into a priority queue. From the priority queue Pop the path with the lowest cost and add it to the set. middle; Repeat the above steps until you obtain... 10 candidate paths.

[0028] Furthermore, the expressions for the delay constraint, jitter constraint, and reliability constraint are as follows:

[0029]

[0030]

[0031] in, Indicates link The time delay, Indicates the first The cost of each path, Indicates the maximum delay. Indicates link Average latency, Indicates the maximum jitter delay. Indicates link The probability of failure, Represents a node The probability of failure, Indicates the lower limit of reliability. Describes the set of links for path P. The set of nodes representing path P.

[0032] Furthermore, the expression for the comprehensive scoring function is as follows:

[0033] in, This represents the comprehensive scoring function. , , These represent the weighting factors for latency constraints, jitter constraints, and reliability constraints, respectively. Indicates the first Total delay of the path, Indicates the first jitter latency of the path, Indicates the first The reliability of the path.

[0034] The beneficial effects of this invention are: This invention provides a low-altitude deterministic routing method based on in-band telemetry and high-order graph learning, implemented using a high-order temporal graph neural network based on in-band telemetry. According to the architecture of data acquisition layer—network modeling layer—routing decision layer, a deep perception, intelligent prediction and highly reliable routing mechanism for low-altitude networks is established to ensure deterministic transmission of low-altitude network services.

[0035] This invention introduces in-band telemetry technology, which can capture dynamic changes in the network in real time in low-altitude communication environments, and realizes fine-grained detection of the entire link, thus providing a higher precision link awareness capability than existing technologies.

[0036] This invention, by introducing a high-order temporal graph neural network, can more efficiently handle high-frequency topological changes in low-altitude networks and better capture complex network interactions and dependencies. Compared to traditional temporal graph neural networks, it offers better scalability and prediction accuracy.

[0037] The low-altitude deterministic routing mechanism proposed in this invention can select the optimal path that meets the requirements of latency, jitter, and reliability based on network state prediction results. Compared with traditional load balancing-based routing methods, this mechanism can better guarantee the high-reliability transmission of critical tasks and is particularly suitable for the special service requirements in low-altitude communication. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0040] Example like Figure 1 As shown, this invention provides a low-altitude deterministic routing method based on in-band telemetry and high-order graph learning, the implementation of which is as follows: S1. At the data acquisition layer, an auxiliary probe mechanism optimized for global link coverage is used to detect the fine-grained paths of the entire low-altitude network and obtain INT data. The implementation method is as follows: S101. Construct a bipartite graph of the low-altitude network topology based on candidate probe nodes and links to be observed. Use Hopcroft-Karp to obtain the maximum matching set of the bipartite graph. Specifically, based on the bipartite graph, use breadth-first search to construct a hierarchical graph and calculate distances to find the length of the shortest augmenting path; based on the length of the shortest augmenting path, use depth-first search to find augmenting paths where vertices do not intersect; update the matching, and repeat until no augmenting path can be found in the breadth-first search, thus obtaining the maximum matching set of the bipartite graph. S102. Based on the maximum matching set, calculate the minimum edge-covered edge set; S103. By using the minimum edge to cover the edge set, INT data is obtained by generating probe message paths. In this process, a method of multiple probe nodes covering different edges simultaneously is adopted to improve the detection speed.

[0041] In this embodiment, a bipartite graph of the low-altitude network topology is constructed based on candidate probe nodes and the links to be observed. The maximum matching set of the bipartite graph is obtained using the Hopcroft-Karp maximum matching method. The Hopcroft-Karp maximum matching method process is as follows: a layered graph is constructed using breadth-first search and distances are calculated to find the length of the shortest augmenting path; a depth-first search is used to find augmenting paths where vertices do not intersect; the matching is updated, and this process is repeated until no augmenting path can be found in the breadth-first search. According to graph theory, the minimum edge cover of a bipartite graph = all vertices - maximum matching. The final calculated edge set is the minimum edge cover, which covers all nodes while minimizing the number of probe edges.

[0042] Probe nodes periodically perform INT probes and report link status information to the INT controller. Probe message paths are generated from the minimum edge coverage set, and multiple probes simultaneously covering different edges are used to improve probe speed. INT probe messages are responsible for collecting three types of data: link quality of service (QoS), node status, and timestamps. Link QoS includes latency, packet loss rate, bandwidth utilization, and queue usage; node status includes node location and node load; and timestamps are used to calculate one-way link latency and synchronization data.

[0043] This invention proposes an auxiliary probe mechanism for global link coverage optimization, which uses INT technology to achieve fine-grained detection of the entire low-altitude network. While ensuring global monitoring of the low-altitude network, it reduces telemetry and processing overhead, providing data support for subsequent modeling and decision-making.

[0044] S2. In the network modeling layer, based on INT data, a network fault prediction mechanism based on a high-order temporal graph neural network is used to calculate the probability of node and link failures. The implementation method is as follows: S201. Preprocess the INT-based data; S202. Based on the preprocessed INT data, the topology of the low-altitude network at different timestamps is represented as a graph sequence; S203. Based on the graph sequence, perform high-order neighbor modeling on the nodes to complete the embedding update of the nodes; S204. Update the link features based on the embedding update results of the two end nodes; S205. Based on the update of link features and node features, the nodes and links of different time slices are stacked to form a time series feature sequence, and then input into a high-order time series graph neural network. S206. Use a high-order temporal graph neural network to predict the future dynamic characteristics of nodes or links; S207. Based on the predicted dynamic characteristics, predict the state at the next moment; S208. Input the predicted state of the next moment into the logistic regression model to calculate the probability of node and link failure.

[0045] In this embodiment, the collected INT data undergoes data preprocessing. Data preprocessing includes data cleaning and feature normalization; data cleaning uses the historical mean method to fill in missing data to ensure the continuity of time-series data; feature normalization avoids some features from dominating the training process, which could affect the model's convergence speed and prediction accuracy.

[0046] The topology of the low-altitude network at different timestamps is represented as a graph sequence expression:

[0047] in, Represents a set of nodes. Indicates time The set of edges, This indicates node characteristics such as location, load, queue length, etc. This represents edge characteristics such as latency and bandwidth utilization. Represents a graph sequence. Indicates time The graph sequence transforms time series data into graph sequences. Higher-order Temporal Graph Networks (HTGNs) can simultaneously capture spatial topology and temporal dynamic information, enabling high-precision prediction of low-altitude networks.

[0048] Higher-order neighbor modeling is performed on the High-Order Temporal Graph Neural Network (HTGN) to capture the nonlocal dependencies of how changes in some links affect the states of other links through multi-hop nodes. The definition of a higher-order neighbor is:

[0049] The node is embedded and updated; the node is at the [number]th [time]. The expression for the higher-order graph convolution of a layer is:

[0050] This invention uses the ReLU function as a non-linear activation function.

[0051] The formula for embedding and updating link features is:

[0052] in, It is a multi-layer perceptron (MLP) used to fuse node embeddings and edge attributes to generate high-order temporal representations.

[0053] The node and link representations from different time slices are input into a recurrent neural network to capture the dynamic evolution of the network state and obtain predictive representations for future times. The higher-order embeddings of nodes and links constitute the time series, specifically expressed as:

[0054] The time series data is input into a gated recurrent unit to predict the future dynamic characteristics of a node or link. The specific expression is as follows:

[0055] Based on the temporal representation of nodes and links, a feedforward layer is used to predict the state at the next time step. The specific expression is as follows:

[0056] The core of Higher-Order Temporal Graph Neural Networks (HTGNs) lies in characterizing the complex interactions among node groups through a higher-order structure, and continuously updating node states by combining memory modules and temporal aggregation mechanisms. Compared to traditional temporal graph neural networks and temporal graph attention networks, HTGNs can more efficiently handle the problem of high-frequency topological changes in low-altitude scenarios, reducing memory and computational overhead.

[0057] Logistic regression is introduced to calculate the probability of node and link failures, specifically expressed as:

[0058]

[0059] in, This represents the probability of a link failure. This represents the probability of node failure. The weights represent the link failure predictions. Indicates the link prediction status. Indicates the predicted state of the node. The weights represent the node failure predictions. and These represent the bias terms for link failure prediction and node failure prediction, respectively. This represents the activation function. The mapping weights represent the link characteristics. Represents the dynamic characteristics of the connection. This indicates the bias in the link feature prediction. Represents the node feature mapping weights. Represents the dynamic characteristics of a node. This indicates the bias in node feature prediction. Indicates that the node is Dynamic characteristics at any given moment The time series representing the higher-order embeddings of nodes. Indicates the link is in Dynamic characteristics at any given moment This represents a time series composed of higher-order embeddings of edges. Represents a node exist Moment embedding, Represents a node exist Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in the th Layer embedding vectors, This represents a multilayer perceptron. Indicates the connection node and nodes The link in the th Layer embedding vectors, Indicates the first Layer nodes Embedded vector, Indicates the first Layer Neighbor weight matrix of order, Indicates the first Layer nodes Embedded vector, No. Layer nodes Embedded vector, Indicates the first The weight of the layer node itself, where K represents the maximum order of its higher-order neighbors. Indicates a higher-order neighbor. Indicates the first 1 node Represents a set of nodes.

[0060] This invention proposes a network fault prediction and routing mechanism based on a high-order temporal graph neural network, which realizes effective modeling of the historical dynamic characteristics of nodes and links, mines higher-order temporal relationships, and achieves accurate prediction of network faults.

[0061] S3. At the routing decision layer, based on the calculated failure probability, the optimal deterministic route is obtained using a low-altitude deterministic routing mechanism that minimizes the failure probability. The implementation method is as follows: S301. The probability of node and link failure is mapped using a negative logarithm and quantified into an additive path cost. The path cost is accumulated segment by segment to obtain the cost map of the low-altitude network. S302. Based on the cost graph, generate multiple candidate paths; S303. Score and filter the candidate paths, and select the paths that meet the time delay constraints, jitter constraints and reliability constraints as feasible paths; S304. Define the comprehensive scoring function; S305. Use the comprehensive scoring function to obtain the optimal deterministic path among the feasible paths.

[0062] In this embodiment, the failure probability is quantified into a weighted cost. A negative logarithmic mapping method is used to transform the path cost into a summation corresponding to a failure-free probability. The specific expression is as follows:

[0063]

[0064]

[0065] in, Indicates the weight of the link. Represents the weight of a node. Representing a path The total cost is thus obtained, resulting in the cost map of the low-altitude network.

[0066] Using the Yen algorithm, multiple candidate paths are generated based on the cost graph. First, Dijkstra's algorithm is used to find the first shortest path. ,Bundle Add to solution set For the first one found Path Choose a fixed prefix, take the last node of the fixed prefix as the new starting point, and use Dijkstra's algorithm again to find a new path on a temporary graph with some edges or nodes removed, then add it to a priority queue. .from Pop up the path with the lowest cost and add it. Repeat the above steps until you obtain... 10 candidate paths.

[0067] Candidate paths are scored and filtered, and paths that meet the constraints of latency, jitter, and reliability are selected as feasible paths. The specific expressions for latency, jitter, and reliability constraints are:

[0068]

[0069]

[0070] in, Indicates link The time delay, Indicates the first The cost of each path, Indicates the maximum delay. Indicates link Average latency, Indicates the maximum jitter delay. Indicates link The probability of failure, Represents a node The probability of failure, Indicates the lower limit of reliability. Describes the set of links for path P. Let P be the set of nodes. Feasible paths are obtained by filtering from this set. For paths that meet the conditions, a comprehensive scoring function is defined:

[0071] in, This represents the comprehensive scoring function. , , These represent the weighting factors for latency constraints, jitter constraints, and reliability constraints, respectively. Indicates the first Total delay of the path, Indicates the first jitter latency of the path, Indicates the first The reliability of the path.

[0072] The weighting factors can be adjusted according to actual needs. The optimal deterministic path is obtained from the feasible paths by the comprehensive scoring function, and the controller issues forwarding entries and resource reservation strategies. If the predicted failures of links and nodes exceed the threshold, the path selection process is re-executed.

[0073] This invention proposes a low-altitude deterministic routing mechanism based on minimizing the probability of failure. By combining network prediction results with differentiated service requirements, it supports deterministic routing decisions that meet constraints on latency, jitter, and reliability, thereby ensuring highly reliable transmission for critical tasks.

[0074] In summary, this paper proposes an auxiliary probe mechanism for low-altitude deterministic routing based on in-band telemetry and high-order graph learning, implemented using a high-order temporal graph neural network based on in-band telemetry. At the data acquisition layer, this mechanism optimizes global link coverage, enabling fine-grained detection of the entire low-altitude network link. While ensuring global monitoring of the low-altitude network, it reduces telemetry and processing overhead, providing data support for subsequent modeling decisions. At the network modeling layer, a network fault prediction mechanism based on a high-order temporal graph neural network is proposed. Compared to traditional graph neural networks, this high-order temporal graph neural network can not only effectively model the historical dynamic characteristics of nodes and links in the low-altitude network but also mine higher-order temporal relationships, achieving accurate prediction of network faults. At the routing decision layer, a low-altitude deterministic routing mechanism based on fault probability minimization is proposed. By combining network prediction results with differentiated service requirements, this mechanism supports deterministic routing decisions that meet latency, jitter, and reliability constraints, thereby ensuring highly reliable transmission of critical tasks.

[0075] In this embodiment, for data acquisition and network status assessment, alternative solutions can employ one or more combinations of methods such as in-band telemetry, out-of-band telemetry, active probes, passive traffic mirroring, or hybrid measurements to obtain network performance indicators. To balance measurement accuracy and bandwidth overhead, telemetry data can be reported after approximate processing such as aggregation and compression. The data can be in raw packet-by-packet format or in aggregated form. The system can also dynamically adjust the sampling frequency and detection coverage based on traffic priority and quality of service level to implement selective or collaborative probe mechanisms, thereby reducing the detection load while ensuring the integrity of observations.

[0076] In the network state modeling and performance prediction phase, alternative solutions can employ high-order temporal graph neural network models, or any prediction mechanism capable of capturing topological and temporal correlations. The prediction unit may include, but is not limited to: spatiotemporal graph Transformer models based on graph attention mechanisms, time-varying graph neural networks, link-level sequence prediction models based on recurrent neural networks, statistical prediction models based on Kalman filtering or Bayesian inference, and hidden Markov models. The prediction output can be a deterministic numerical value or a probability distribution containing uncertainty measures, and can be combined with external environmental features for multi-source fusion to improve the reliability and robustness of the prediction.

[0077] At the routing decision and scheduling level, the deterministic path selection module of the alternative can be implemented using arbitrary optimization, heuristic, or learning algorithms based on the prediction results. The path selection process can be carried out through multi-constraint shortest path algorithms, linear programming, integer programming, model predictive control, or rolling optimization methods; it can also adaptively learn the optimal path selection strategy through reinforcement learning, deep reinforcement learning, or imitation learning, etc., which can support single-path or multi-path traffic allocation and dynamically adjust the routing strategy according to the prediction confidence.

Claims

1. A low-altitude deterministic routing method based on in-band telemetry and high-order graph learning, characterized in that, Includes the following steps: S1. At the data acquisition layer, an auxiliary probe mechanism optimized for global link coverage is used to detect the fine-grained path of the entire link of the low-altitude network and obtain INT data. S2. In the network modeling layer, based on INT data, the probability of node and link failures is calculated using a network fault prediction mechanism based on a high-order temporal graph neural network. S3. At the routing decision layer, based on the calculated failure probability, the optimal deterministic path is obtained using a low-altitude deterministic routing mechanism that minimizes the failure probability.

2. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 1, characterized in that, S1 includes the following steps: S101. Construct a bipartite graph of the low-altitude network topology based on candidate probe nodes and links to be observed, and use Hopcroft-Karp to obtain the maximum matching set of the bipartite graph. S102. Based on the maximum matching set, calculate the minimum edge-covered edge set; S103. By using the minimum edge to cover the edge set, INT data is obtained by generating probe message paths. In this process, a method of multiple probe nodes covering different edges simultaneously is adopted to improve the detection speed.

3. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 2, characterized in that, The method of using Hopcroft-Karp to find the maximum matching set of a bipartite graph is as follows: Based on the bipartite graph, a layered graph is constructed using breadth-first search and distances are calculated to find the length of the shortest augmenting path. Based on the length of the shortest augmenting path, a depth-first search is used to find augmenting paths where vertices do not intersect. Update the matching and loop until no augmenting path can be found in the breadth-first search, thus obtaining the maximum matching set of the bipartite graph.

4. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 1, characterized in that, S2 includes the following steps: S201. Preprocess the INT-based data; S202. Based on the preprocessed INT data, the topology of the low-altitude network at different timestamps is represented as a graph sequence; S203. Based on the graph sequence, perform high-order neighbor modeling on the nodes to complete the embedding update of the nodes; S204. Update the link features based on the embedding update results of the two end nodes; S205. Based on the update of link features and node features, the nodes and links of different time slices are stacked to form a time series feature sequence, and then input into a high-order time series graph neural network. S206. Use a high-order temporal graph neural network to predict the future dynamic characteristics of nodes or links; S207. Based on the predicted dynamic characteristics, predict the state at the next moment; S208. Input the predicted state of the next moment into the logistic regression model to calculate the probability of node and link failure.

5. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 4, characterized in that, The expressions for the probabilities of node and link failures are as follows: in, This represents the probability of a link failure. This represents the probability of node failure. The weights represent the link failure predictions. Indicates the link prediction status. Indicates the predicted state of the node. The weights represent the node failure predictions. and These represent the bias terms for link failure prediction and node failure prediction, respectively. For activation function, The mapping weights represent the link characteristics. Represents the dynamic characteristics of the connection. This indicates the bias in the link feature prediction. Represents the node feature mapping weights. Represents the dynamic characteristics of a node. This indicates the bias in node feature prediction. Indicates that the node is Dynamic characteristics at any given moment The time series representing the higher-order embeddings of nodes. Indicates the link is in Dynamic characteristics at any given moment This represents a time series composed of higher-order embeddings of edges. Represents a node exist Moment embedding, Represents a node exist Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in Moment embedding, Indicates the connection node and nodes The link in the th Layer embedding vectors, This represents a multilayer perceptron. Indicates the connection node and nodes The link in the th Layer embedding vectors, Indicates the first Layer nodes Embedded vector, Indicates the first Layer Neighbor weight matrix of order, Indicates the first Layer nodes Embedded vector, No. Layer nodes Embedded vector, Indicates the first The weight of the layer node itself, where K represents the maximum order of its higher-order neighbors. Indicates a higher-order neighbor. Indicates the first 1 node Represents a set of nodes.

6. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 1, characterized in that, S3 includes the following steps: S301. The probability of node and link failure is mapped using a negative logarithm and quantified into an additive path cost. The path cost is accumulated segment by segment to obtain the cost map of the low-altitude network. S302. Based on the cost graph, generate multiple candidate paths; S303. Score and filter the candidate paths, and select the paths that meet the time delay constraints, jitter constraints and reliability constraints as feasible paths; S304. Define the comprehensive scoring function; S305. Use the comprehensive scoring function to obtain the optimal deterministic path among the feasible paths.

7. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 6, characterized in that, S302 includes the following steps: Find the first shortest path and the shortest path Add to solution set middle; For the found Path Choose a fixed prefix, take the last node of the fixed prefix as the new starting point, find a new path again on the temporary graph after removing some edges or nodes, and put it into a priority queue. From the priority queue Pop the path with the lowest cost and add it to the set. middle; Repeat the above steps until you obtain... 10 candidate paths.

8. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 6, characterized in that, The expressions for the delay constraint, jitter constraint, and reliability constraint are as follows: in, Indicates link The time delay, Indicates the first The cost of each path, Indicates the maximum delay. Indicates link Average latency, Indicates the maximum jitter delay. Indicates link The probability of failure, Represents a node The probability of failure, Indicates the lower limit of reliability. Describes the set of links for path P. The set of nodes representing path P.

9. The low-altitude deterministic routing method based on in-band telemetry and high-order graph learning according to claim 6, characterized in that, The expression for the comprehensive scoring function is as follows: in, This represents the comprehensive scoring function. , , These represent the weighting factors for latency constraints, jitter constraints, and reliability constraints, respectively. Indicates the first Total delay of the path, Indicates the first jitter latency of the path, Indicates the first The reliability of the path.

Images