Link stability and energy awareness-based low-altitude intelligent networking routing optimization method

Abstract

The invention relates to the technical field of low-altitude intelligent networking optimization, and discloses a low-altitude intelligent networking routing optimization method based on link stability and energy perception. The method comprises the steps that a node collects multi-dimensional state information including position, energy and signal intensity; predicting and calibrating a link outage probability based on the bidirectional long short-term memory network; defining an interruption risk regulation factor and a comprehensive path weight model by combining the interruption probability, the node energy and the link quality; accordingly, multiple relay nodes are elected, and self-adaptive adjustment and coverage detection are achieved; calculating path cost, selecting a main path and a backup path, and triggering pre-exchange when the interruption risk exceeds a threshold value; and maintaining double forwarding tables to realize load balancing data forwarding, and dynamically controlling a message updating period. According to the method, the routing reliability, the energy efficiency and the self-adaptive maintenance capability in the low-altitude dynamic network are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of low-altitude intelligent network route optimization technology, specifically a low-altitude intelligent network route optimization method based on link stability and energy perception. Background Technology

[0002] Traditional static link quality assessment methods are ineffective in handling the dynamically changing environment of Flying Ad Hoc Networks (FANETs), especially under conditions of high-speed movement of low-altitude communication nodes and complex weather (such as wind speed and heading). In this scenario, the probability of link outages is difficult to predict accurately, leading to delayed routing and frequent link interruptions. Therefore, this patent introduces a link outage probability prediction mechanism based on Bidirectional Long Short-Term Memory (Bi-LSTM) networks. By combining historical data with real-time network status and considering the impact of multi-dimensional environmental factors on link outages, it accurately identifies link outage risks in advance, enabling dynamic prediction and timely adjustments, thereby improving network stability, adaptability, and fault tolerance.

[0003] Traditional MRP (Multi-Point Relay) node election mechanisms primarily rely on adjacency relationships and coverage requirements. However, in highly dynamic environments such as low-altitude communication, this mechanism fails to effectively consider factors such as link stability and node energy. This can lead to network topology changes and quality degradation due to relay node energy depletion or declining quality of adjacent links, resulting in low routing efficiency or even communication interruptions. Therefore, this patent proposes a multi-dimensional MRP node election and stability assurance mechanism. By comprehensively considering multiple factors such as link interruption risk, node remaining energy, and network coverage, it prioritizes relay nodes with higher stability and sufficient energy, ensuring reliable long-term network operation in highly dynamic environments and avoiding communication interruptions caused by energy issues or link instability.

[0004] In traditional routing protocols, the periodic broadcast mechanism of control signaling is ill-suited to highly dynamic environments with frequent topology changes, leading to excessive control overhead. This is particularly problematic in networks with large-scale low-altitude UAV deployments and high-speed movement, where increased communication burden and reduced network efficiency occur. Furthermore, fixed-period signaling updates struggle to respond promptly to changes in link status. Therefore, this patent addresses this issue by employing a risk-aware adaptive maintenance and control overhead optimization mechanism. This mechanism extends the control signaling cycle when the network topology is stable and shortens it when links are unstable or topology changes are frequent. This reduces unnecessary control signaling, optimizes the utilization efficiency of limited resources in low-altitude networks, ensures real-time information updates, and improves overall network efficiency and stability. Summary of the Invention

[0005] The purpose of this invention is to provide a low-altitude intelligent network routing optimization method based on link stability and energy awareness, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a low-altitude intelligent network routing optimization method based on link stability and energy awareness, the method comprising: Multidimensional status information is collected by nodes in the low-altitude intelligent network. The multidimensional status information includes node identifier, timestamp, three-dimensional coordinate position, remaining energy percentage, received signal strength indication, number of neighboring nodes, wind speed, heading angle, node status, packet loss rate, and link quality. Based on the multidimensional state information, a bidirectional long short-term memory network model is used to predict the link interruption probability, and the predicted value is subjected to probability calibration and smoothing to generate a calibrated link interruption probability. The link survival probability is calculated based on the calibrated link outage probability. An outage risk adjustment factor and a comprehensive path weight model are defined by combining the node's remaining energy and link quality. The score of each candidate node is calculated based on the outage risk adjustment factor, node coverage, and remaining energy. Multi-point relay nodes are elected based on the scores, and adaptive adjustment and coverage detection are achieved. The path cost is calculated based on the comprehensive path weight model, and the shortest path algorithm is used to select the primary path and backup path. When the calibrated link interruption probability exceeds a preset threshold, a pre-switching strategy is triggered. Maintain dual forwarding tables for primary and backup routes to achieve data forwarding and load balancing based on node load; dynamically adjust the update cycle of neighbor discovery messages and topology control messages according to the calibrated link outage probability.

[0007] Preferably, the step of collecting multi-dimensional status information through nodes in the low-altitude intelligent network includes: Nodes periodically broadcast neighbor discovery messages, which include the node's unique identifier, current timestamp, three-dimensional coordinates, remaining energy percentage, and a description of communication capabilities. After receiving the neighbor discovery message, other nodes update their local neighbor table and calculate the link packet loss rate and received signal strength indication. The receiving node integrates the wind speed, heading angle, and node status information it senses to construct a neighbor table containing all multi-dimensional status information.

[0008] Preferably, the step of predicting the link interruption probability using a bidirectional long short-term memory network model based on multidimensional state information includes: The multidimensional state information is scaled to generate an input vector that includes velocity magnitude, acceleration magnitude, angle vector, coordinates, energy percentage, link quality index, neighbor density, wind speed, and time interval. The input vector is fed into a pre-trained bidirectional long short-term memory network model, which is pre-trained using historical low-altitude network operation data. The original link interruption probability prediction value is output through forward and backward time-series processing. The original link interruption probability prediction value is probabilistically calibrated using the Platt scaling method and smoothed using an exponential moving average. The smoothing factor α of the exponential moving average is a configurable parameter with a value range of (0,1) and a default value of 0.2.

[0009] Preferably, the step of calculating the link survival probability based on the calibrated link outage probability includes: The link survival probability is defined as 1 minus the calibrated link failure probability; Calculate the average survival probability of adjacent edges, and combine a small constant to avoid division by zero error. Define the interruption risk adjustment factor as the link survival probability divided by the sum of the average survival probability and the constant. A comprehensive path weight model is constructed by combining link quality indicators and node remaining energy. The comprehensive path weight model is constructed using the integrated link quality index and the remaining energy of the nodes, including: Define a link quality penalty, which is 1 minus the link quality metric, such that the worse the link quality, the larger the metric value. Define a penalty for the remaining energy of a node, which is 1 minus the percentage of the node's remaining energy, such that the lower the node's energy, the larger this value is; The link cost is defined as the linear sum of a risk cost, an energy penalty, and a quality penalty. The risk cost is the interruption risk adjustment factor multiplied by a first weighting coefficient. The energy penalty is the average of the remaining energy penalties of the nodes at both ends of the link multiplied by a second weighting coefficient. The quality penalty is the link quality penalty multiplied by a third weighting coefficient. The first, second, and third weighting coefficients are real numbers greater than zero. Their specific values ​​are preset through offline simulation or based on prior knowledge of network deployment, and can be slowly and adaptively adjusted based on the overall network status. Define the path cost of any candidate path, where the path cost is the sum of the link costs of all links on that path; The reciprocal of the path cost is defined as the path comprehensive weight. The higher the path comprehensive weight, the better the path.

[0010] Preferably, the calculation of the score for each candidate node based on the interruption risk adjustment factor, node coverage, and remaining energy includes: The number of two-hop neighbors covered by a candidate node is used as a coverage metric. Obtain the remaining energy of the candidate nodes; Calculate the average outage risk adjustment factor between the candidate node and its neighboring nodes; Candidate node scores are obtained by weighted summation of coverage, remaining energy, and average interruption risk adjustment factor.

[0011] Preferably, the step of electing multi-point relay nodes based on scores and implementing adaptive adjustment and coverage detection includes: Select the candidate node with the highest score as the multi-point relay node; The coverage of two-hop neighbors is checked periodically. The period of the periodic check can be dynamically set according to the average network mobility or topology change rate. When the coverage is insufficient, the relay node set is dynamically adjusted. Verify relay node paths to avoid loop formation; The selection of the candidate node with the highest score as the multi-point relay node includes: In the set of one-hop neighbors of the current node, sort them according to the scores calculated by the candidate node scoring function; Select the node with the highest score from the sorting results as the preferred multi-point relay node for this node; The identifier and score of the preferred multipoint relay node are broadcast to neighbors through relay node election messages. If a node receives election messages from multiple neighbors all electing it as the preferred relay node, then the node confirms itself as a multipoint relay node and broadcasts a relay announcement message to the entire network. If a node receives multiple relay announcement messages, it determines its own assigned relay node based on the rating of the announcement nodes, thus completing the establishment of a multi-point relay topology.

[0012] Preferably, the step of calculating the path cost based on the comprehensive path weight model includes: The risk cost term is defined based on the calibrated link outage probability. Define energy penalty terms based on node remaining energy penalties; Define quality penalty items based on link quality penalties; The link cost is obtained by adding the risk cost, energy penalty, and quality penalty. The path cost is the sum of all link costs.

[0013] Preferably, the step of using the shortest path algorithm to select the primary path and backup path includes: Apply Dijkstra's algorithm to find the main path with the minimum path cost; At the same time, several backup paths are built as backups; When the probability of link interruption after calibration of the primary path exceeds the threshold, the pre-switching strategy is triggered to switch to the backup path; Preferably, the application of Dijkstra's algorithm to find the main path with the minimum path cost includes: Starting from the source node, initialize its path cost to zero, and set the initial path cost of all other nodes in the network to infinity. Create a set of nodes for which the shortest path has not yet been determined, initially containing all nodes in the network; In each iteration, the node with the lowest known path cost is selected from the set as the current processing node; Iterate through all neighboring nodes of the current processing node. For each neighboring node, calculate the total path cost from the source node to that neighboring node via the current processing node. This total path cost is equal to the known path cost of the current processing node plus the link cost connecting the two nodes. If the calculated total path cost is less than the known path cost of the neighboring node, then update the path cost of the neighboring node to the total path cost, and record the currently processed node as its predecessor node. Remove the currently processed node from the set of nodes with undetermined shortest paths; Repeat the above iterative process until the target node is removed from the set. At this point, based on the predecessor node information recorded by each node, backtrack from the target node to the source node to obtain the main path with the minimum path cost.

[0014] Preferably, the maintenance of the dual forwarding table for the primary and backup routes includes: When the main router is functioning normally, data is forwarded through the main forwarding table. When the pre-swapping policy is triggered, the forwarding table entries of the backup path are sent to the target node in advance. The advance sending operation is executed immediately after the pre-swapping policy is triggered. Load balancing is achieved by adjusting forwarding priorities based on node remaining energy and queue length. Specifically: The method of adjusting forwarding priority based on node remaining energy and queue length to achieve load balancing includes: A load balancing weight is defined for each node. The load balancing weight is calculated by weighting the node's remaining energy percentage and queue length percentage, where the queue length percentage is the ratio of the current data queue length to the maximum queue capacity. When a node forwards a data packet through the primary path or backup path, if there are multiple relay nodes available for the next hop, the node is selected based on the load balancing weight value currently broadcast by each relay node, with the node having a higher weight value having a greater probability of being selected. Nodes periodically calculate and broadcast their latest load balancing weights to achieve dynamic traffic distribution based on real-time load status.

[0015] Preferably, the method further includes: Collect network operation data, and update the parameters of the bidirectional long short-term memory network model and adjust the preset threshold accordingly; specifically including: Record link interruption events, handover latency, packet loss rate, and energy consumption data; The bidirectional long short-term memory network model is fine-tuned online using recorded data. The online fine-tuning is triggered periodically or when the model's prediction error continues to exceed a set threshold. The link interruption probability threshold and control signaling cycle threshold are adaptively adjusted based on long-term statistical results, including path pre-switching success rate, average handover delay, and control signaling overhead ratio.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) Improve the accuracy and timeliness of link interruption prediction By introducing a link interruption probability prediction mechanism based on a bidirectional long short-term memory network model, and fully considering historical information on link status and real-time environmental factors (such as wind speed, heading, node energy, etc.), dynamic and accurate link interruption probability prediction is achieved. Potential link failure risks can be identified in advance, and path selection can be adjusted in a timely manner, thereby improving the dynamic adaptability and fault tolerance of the low-altitude network.

[0017] (2) Optimize MRP node election to improve network stability and energy efficiency. The multi-dimensional MRP node election and stability guarantee mechanism proposed in this invention comprehensively considers multiple factors such as link stability, energy status and network coverage requirements, and prioritizes nodes with high stability and sufficient energy as relays to avoid communication interruptions caused by energy depletion or link instability. This ensures the long-term reliable operation of the low-altitude intelligent network in a highly dynamic environment and improves the fault tolerance and communication efficiency of the low-altitude network.

[0018] (3) Reduce the communication overhead of control signaling and improve resource utilization. By employing a risk-aware adaptive maintenance and control overhead optimization mechanism, this invention can dynamically adjust the update cycle of control signaling, extending the signaling cycle when the network topology is stable and shortening the cycle when topology changes frequently or links are unstable. This reduces unnecessary control message transmission, lowers control overhead, improves network resource utilization, ensures real-time updates of routing information, and guarantees the efficiency and stability of the low-altitude network. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the low-altitude intelligent network routing optimization method based on link stability and energy sensing described in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides a routing optimization method for low-altitude intelligent networks based on link stability and energy awareness. The overall implementation scheme is as follows: Nodes in the low-altitude intelligent network periodically collect and interact with multi-dimensional state information. This multi-dimensional state information includes at least node identifier, timestamp, three-dimensional coordinate position, remaining energy percentage, received signal strength indication, number of neighboring nodes, wind speed, heading angle, node status, packet loss rate, and link quality. The collected information is used to construct and update the node's local neighbor table. Subsequently, based on the historical and current multi-dimensional state information sequence, a pre-trained bidirectional long short-term memory network model is used to predict the original outage probability of each link in the future. This prediction value is probability-calibrated using the Platt scaling method and then smoothed using exponential moving average to finally generate the calibrated link outage probability. The link survival probability is calculated based on the calibrated link outage probability, and an outage risk adjustment factor is further calculated by combining network topology information. Simultaneously, a comprehensive path weight model is constructed using the node's remaining energy percentage and link quality indicators, together with the outage risk adjustment factor. This model is used to quantify the merits of each link and path. In the topology organization phase, each candidate node is scored based on an outage risk adjustment factor, node coverage, and remaining node energy. Multiple relay nodes are elected based on these scores, enabling adaptive adjustment of the relay node set and network coverage detection. In the path selection phase, the primary path and several backup paths are calculated using a shortest path algorithm (such as Dijkstra's algorithm) based on the path cost calculated by the comprehensive path weight model. A preset threshold for link outage probability is established. When the calibrated link outage probability of any link on the primary path exceeds this threshold, a pre-switching strategy is triggered, switching the data flow to the backup path in advance. At the data forwarding level, nodes maintain dual forwarding tables for primary and backup routes. Data is forwarded according to the route status, and traffic allocation is dynamically adjusted based on real-time node load (such as remaining energy and queue length) to achieve load balancing. Furthermore, the entire method includes an adaptive maintenance mechanism that dynamically adjusts the update cycle of neighbor discovery messages and topology control messages based on the calibrated link outage probability. Network performance data is collected over long-term operation to update the parameters of the prediction model and adjust relevant thresholds.

[0022] Example 1: In practical implementation, nodes in the low-altitude intelligent network periodically broadcast neighbor discovery messages. The message content encapsulates the node's multi-dimensional state information, including its unique identifier, current timestamp, three-dimensional coordinates, remaining energy percentage, and a description of its communication capabilities. Node A at timestamp... Its three-dimensional coordinates Remaining energy percentage The communication capability description is encapsulated in the neighbor discovery message and broadcast. Node B successfully receives the broadcast message from node A within its communication range. Node B parses the message content and obtains node A's timestamp. 3D coordinates and percentage of remaining energy Node B calculates the received signal strength indication from Node A based on the received signal power. Node B in the time window The link packet loss rate to node A is calculated by comparing the ratio of the number of neighbor discovery messages received from node A to the expected number of messages. Node B will store the identifier and timestamp of node A. 3D coordinates Remaining energy percentage Received signal strength indication and the calculated packet loss rate The data is recorded in node B's local neighbor table, becoming a record about neighboring node A. Simultaneously, node B integrates the real-time wind speed data acquired by its own sensors. Heading angle and the working status of node B itself. This information is also associated with the current communication context, thereby constructing a multi-dimensional set of state information locally on node B, including node identifier, timestamp, three-dimensional coordinate position, remaining energy percentage, received signal strength indication, number of neighboring nodes, wind speed, heading angle, node status, packet loss rate, and link quality.

[0023] In some embodiments, after recording multidimensional state information, node B continues to preprocess the information to generate an input vector suitable for a bidirectional long short-term memory network model. Node B reads its own timestamp. Historical three-dimensional coordinates With current timestamp 3D coordinates By calculating the coordinate difference and the time difference Get node B in plane and The velocity components along the axial direction are then combined to synthesize the instantaneous velocity magnitude. Based on the velocity sequence of multiple consecutive historical moments, node B calculates the instantaneous acceleration magnitude. Node B will set its own heading angle. Convert to unit direction vector For the three-dimensional coordinates of node A Node B is linearly scaled using preset maximum and minimum values ​​of the network region to obtain the scaled coordinates. The percentage of remaining energy at node A. Scaled to The interval is used to obtain the percentage of remaining energy. Node B is based on the received signal strength indication. and packet loss rate Calculate a comprehensive link quality metric and link quality indicators Scaling is performed. Node B counts the total number of one-hop neighbor nodes recorded in its local neighbor table, which is used as the neighbor density. Wind speed sensed by node B Maximum wind speed scaling applied. Time interval. This is calculated as the difference between the current processing time and the previous model input time. Finally, for the link between node A and node B, the generated input vector... It includes the following components: velocity magnitude Magnitude of acceleration Heading angle unit vector Scaling coordinates of node A Energy percentage of node A Scaling-up link quality metrics Neighbor density Scaling up the wind speed and time interval This input vector It characterizes the dynamic state and link characteristics at one end of the link at a specific moment.

[0024] It is understandable that the generated input vector The input is fed into a pre-trained bidirectional long short-term memory network model. The model receives an input vector consisting of multiple consecutive time steps. The sequence is constructed by the forward layer of the bidirectional long short-term memory (BSSM) network model, which processes the sequence in chronological order, while the backward layer processes the same sequence in reverse chronological order. The BSSM network model learns long-term dependencies in the sequence through its internal gating mechanism, ultimately generating a scalar value at the output layer. scalar value This is the original predicted probability of a link break between node A and node B in the next time interval, as predicted by the bidirectional long short-term memory network model. The range of values ​​is arrive Between. Original predicted probability It is the direct output of the uncalibrated model.

[0025] Optionally, to improve the original prediction probability To improve reliability, the Platt scaling method is used to adjust the original predicted probabilities. Probabilistic calibration is performed. The Platt scaling method utilizes a preserved, independent validation dataset containing historical input vector sequences and their corresponding actual link interruption labels. or The Platt scaling method treats the calibration process as a logistic regression problem, fitting two parameters. and , the original predicted probability Mapped to calibrated probabilities The mapping relationship can be represented as: parameter and The calibrated probability is obtained by minimizing the negative log-likelihood loss on the validation dataset. It is closer to the true posterior probability of the interruption in a statistical sense.

[0026] In some embodiments, to reduce the probability after calibration Short-term fluctuations affect the calibrated probability. An exponential moving average is performed to obtain a smoothed probability value for subsequent routing decisions. The exponential moving average process requires a smoothing factor. , Values ​​in arrive Between. The smoothing process is defined as: in Indicates the current time The smoothed link interruption probability, Indicates the current time The calibrated probabilities output by the Platt scaling method. Indicates the previous moment The smoothed link interruption probability. Smoothing factor. Controlling new observations Smoothing values ​​with historical values The weights between them, the larger ones This reduces the probability of link interruption after smoothing. More sensitive to recent changes, smaller This reduces the probability of link interruption after smoothing. More stable. The smoothed link outage probability obtained after exponential moving average processing. This is the calibrated link outage probability used in subsequent steps.

[0027] Example 2: In practice, nodes calculate the link survival probability based on the link failure probability value output from the bidirectional long short-term memory network model and after calibration and smoothing. For any connected node... With nodes wireless link The calibrated link interruption probability is denoted as ,link Link survival probability Defined as minus ,Right now: To evaluate the link To determine the relative stability within its local environment, an interruption risk adjustment factor needs to be calculated. The calculation of the interruption risk adjustment factor involves the link... The set of adjacent edges between two nodes, for each node Get the set of all its neighbor nodes. Calculate the set Each link (i.e., node) With any of his neighbors Links between Link survival probability The average value is denoted as Similarly, compute nodes The average survival probability of all adjacent links .link Average survival probability of adjacent edges Defined as and The average value, that is:

[0028] Introducing a very small positive constant To prevent the denominator from being zero, thus interrupting the risk adjustment factor. Defined as a link Link survival probability Divide by the average survival probability of adjacent edges With constant The sum, expressed as: Disruption risk adjustment factor This reflects the link under the current local network topology. Compared to the stability of its surrounding links, the interruption risk adjustment factor A larger value indicates a higher link The more stable it is.

[0029] In some embodiments, the construction of a comprehensive path weight model requires integrating link stability, node energy state, and link transmission quality. A link quality penalty is defined, and its calculation is based on the original link quality metrics. Original link quality indicators It is a comprehensive function of parameters such as received signal strength indication and packet loss rate, and its value range is usually within... arrive Between these values, a larger value indicates better link quality. To impose a larger penalty on links with lower quality in the cost model, a link quality penalty is defined as follows:

[0030] Therefore, link quality metrics A higher value indicates better link quality. Defines the remaining energy of a node. The percentage of its remaining energy is denoted as The remaining energy of a node is the ratio of its remaining energy to its original total energy. The remaining energy penalty is defined as follows:

[0031] Therefore, residual energy penalty The larger the value, the stronger the node. The lower the remaining energy. Based on the interruption risk adjustment factor. Residual energy penalty and and link quality penalties Build links Link cost Link cost It consists of three linearly added terms: the risk cost term is the interruption risk adjustment factor. The reciprocal of the first weighting coefficient The energy penalty term is the average of the remaining energy penalties at both ends of the link. Multiply by the second weighting factor The quality penalty item is the link quality penalty. Multiply by the third weighting factor Link cost The calculation formula is expressed as follows: in, , , These are preset non-negative weighting coefficients used to adjust the impact of different factors on link cost. The relative importance of the source node in the network. to the destination node Candidate paths ,path A path consists of a series of sequentially connected links. Path cost Defined as a path Link cost of all links The sum is:

[0032] Path comprehensive weight Defined as path cost The reciprocal of, that is: Path comprehensive weight The higher the value, the better the path performance in terms of stability, energy balance, and transmission quality. The better.

[0033] It's understandable that in the process of electing multi-point relay nodes, a score needs to be calculated for each candidate node. For a candidate node... Node coverage index Defined as a candidate node The number of two-hop neighbor nodes that can be covered, i.e., through candidate nodes The set of all neighboring nodes of a neighboring node that can communicate directly (excluding duplicates and candidate nodes). Size of itself. Retrieve candidate nodes. Remaining energy Calculate candidate nodes The average of the interruption risk adjustment factors among all its neighboring nodes is denoted as , Reflects candidate nodes The average relative stability of links between the candidate node and all its neighbors. rating Coverage index by weighted summation Remaining energy and average interruption risk adjustment factor Obtained. Specifically: in , , These are the weight coefficients for the corresponding terms. Since... A larger value indicates that the candidate node with higher remaining energy contributes more. The result of the weighted summation... This is the final score used for relay node election. Higher-ranking candidate nodes are considered more suitable to serve as multi-point relay nodes.

[0034] Optional, first weighting coefficient Second weighting coefficient Third weighting coefficient and scoring weighting coefficients , , The specific values ​​can be preset through offline simulation or based on prior knowledge of network deployment. In actual implementation, these coefficients can be fixed or designed as parameters that can be slowly and adaptively adjusted according to the overall network status (such as average node energy and average link stability) to adapt to network needs at different stages. Interruption risk adjustment factor Small constants introduced in the calculation It is a much smaller survival probability than the average survival probability of a typical adjacent edge. Positive numbers, for example Its function is to ensure that the denominator is not zero, thus preventing calculation errors. Link quality metrics The calculation ensures the consistency of the cost model, meaning that all indicators that negatively impact path selection (poor link quality, low node energy, relatively unstable links) are transformed into positively correlated penalty terms, thus reducing path cost. The increase clearly implies a path-wide weighting The decrease.

[0035] Example 3: In practice, the election process for multi-point relay nodes begins with each node independently evaluating its one-hop neighbor set. Each node ranks all candidate nodes in its one-hop neighbor set based on a candidate node score calculated using an interruption risk adjustment factor, node coverage, and remaining energy. Each node maintains a local list recording the identifier of each neighbor node and its corresponding candidate node score. The ranking operation is performed in descending order of candidate node scores, generating an ordered sequence of candidate nodes. From the ranking results, the node selects the neighbor node with the highest candidate node score as its preferred multi-point relay node. The node generates a relay node election message, the payload of which includes the node's identifier, the identifier of the selected preferred multi-point relay node, and the candidate node score of that preferred multi-point relay node. The node broadcasts the relay node election message within its communication range, ensuring that all one-hop neighbor nodes receive the message.

[0036] In some embodiments, nodes in a network receive relay node election messages from multiple neighboring nodes during operation. Each node parses each received relay node election message and checks if the identifier of the preferred multipoint relay node specified in the message matches its own identifier. If a node finds that multiple messages from different neighboring nodes list its own identifier as the preferred multipoint relay node, the node confirms that it has been selected as a relay node by multiple neighbors. This node then assumes the role of a multipoint relay node and broadcasts a relay announcement message to the entire network. The relay announcement message includes the multipoint relay node's own identifier, its candidate node score, and an optional summary of the list of neighboring nodes that have selected the multipoint relay node as a relay. The broadcast of the relay announcement message aims to announce the existence of the multipoint relay node and its service area.

[0037] It's understandable that a regular node might be within the communication coverage of multiple multi-relay nodes simultaneously, thus potentially receiving relay advertisement messages from multiple different multi-relay nodes. The node needs to make a home decision among these multi-relay nodes to complete topology construction. Each node maintains a list of received relay advertisement messages, recording the identifier of each multi-relay node and its candidate node score carried in the advertisement. The node's home decision rule is to select the multi-relay node with the highest candidate node score as its home relay node. The decision formula is expressed as:

[0038] in The representative node selects its own relay node. The set of multi-point relay nodes corresponding to all relay announcement messages received by the representative node. Represents a multipoint relay node in the set. The candidate nodes are scored. A node formally joins the cluster centered on the selected home relay node by sending an association request message to the selected home relay node and receiving an association confirmation message, thus completing the establishment of the multi-point relay topology.

[0039] Optionally, after the multi-point relay topology is established, coverage checks and maintenance need to be performed periodically. Each multi-point relay node periodically checks the status of its two-hop neighbors that it should cover. Based on its list of associated member nodes, the multi-point relay node derives the theoretically coverable set of two-hop neighbor nodes. The multi-point relay node obtains the actual set of communicable two-hop neighbor nodes by listening to network communications or receiving specific probe feedback. The multi-point relay node compares the theoretical set with the actual set. If it finds that nodes in the actual set that are not covered by the theoretical set for a long time, or newly emerging two-hop neighbor nodes that are not covered by any relay, it determines that there is insufficient coverage. Once insufficient coverage is detected, the multi-point relay node with insufficient coverage or its related member nodes can initiate a local or global relay node re-election process. The re-election process may include adjusting scoring weights, rebroadcasting relay node election messages, or triggering new relay announcements, thereby dynamically adjusting the set of relay nodes in the network to ensure complete topology coverage.

[0040] In practice, the path construction process must verify paths passing through relay nodes to avoid loops. When a source node initiates route discovery or a node calculates a path to a destination node, the path may pass through one or more multi-point relay nodes. When constructing forwarding table entries for such paths, nodes need to check the node identifiers in the path sequence. The verification logic ensures that no node identifier appears twice or more in the path node identifier sequence from the source node to the destination node. For example, the node sequence of a candidate path is... The verification algorithm will sequentially check the identifiers. Ensure all identifiers are unique. If duplicate identifiers are found in the sequence, such as... If a loop is detected, the path is determined to be a loop, and the path will be discarded or the route discovery mechanism will be required to recalculate it. Loop verification is a necessary check step before a path is finally adopted and added to the forwarding table, preventing packets from looping infinitely in the network.

[0041] Example 4: In its implementation, Dijkstra's algorithm is used to find the main path with the minimum path cost from the source node to the target node. Dijkstra's algorithm begins at the source node, initializing its known path cost to zero. The known path costs of all other nodes in the network topology are initialized to an extremely large value to represent infinity. Dijkstra's algorithm maintains a set of nodes with undetermined shortest paths, initially containing all nodes in the network. Dijkstra's algorithm enters an iterative loop, selecting the node with the minimum known path cost from the set of undetermined shortest path nodes as the current processing node in each iteration. Dijkstra's algorithm traverses all neighbor nodes of the current processing node. For each neighbor node, it calculates the temporary path cost from the source node to that neighbor node via the current processing node. The temporary path cost equals the known path cost of the current processing node plus the link cost connecting the current processing node and that neighbor node. If the calculated temporary path cost is less than the known path cost currently stored in the neighbor node, the known path cost of the neighbor node is updated to the temporary path cost, and the current processing node is recorded as the predecessor node of that neighbor node. The currently processed node is removed from the set of nodes with an undetermined shortest path. Dijkstra's algorithm repeats the above steps of selection, traversal, calculation, update, and removal until the target node is removed from the set of nodes with an undetermined shortest path. At this point, starting from the target node, the algorithm backtracks to the source node based on the predecessor node information recorded by each node. The resulting node sequence is the main path with the minimum path cost.

[0042] In some embodiments, several backup paths need to be constructed as backups while calculating the primary path. The construction of backup paths can be achieved by modifying the Dijkstra algorithm to record suboptimal paths. During the process of executing the Dijkstra algorithm to find the primary path, not only is the minimum path cost and predecessor node recorded for each node, but also the suboptimal path cost and corresponding predecessor node information for each node. After the Dijkstra algorithm completes the primary path search, a backup path with the second smallest path cost, which is not exactly the same as the primary path, can be obtained by backtracking the suboptimal predecessor information. Another method for constructing backup paths is to apply... Shortest path algorithm The shortest path algorithm can systematically find the path with the minimum cost from the source node to the target node. A simple path, in which It is a preset positive integer. The shortest path algorithm can be implemented based on a variant of Dijkstra's algorithm or a recursive method that deletes links and nodes. The final output is an ordered list of paths, which contains the main path with the minimum path cost and several backup paths with progressively increasing path costs.

[0043] It is understandable that the continuous availability of the main path needs to be ensured by monitoring the link status, and a preset threshold for the probability of link failure should be established. Each node periodically acquires the calibrated link outage probability of each link on the main path. When the calibrated link outage probability of any link on the main path exceeds a preset threshold, the node will be terminated. When a data packet arrives, a pre-switching strategy is triggered. The core operation of the pre-switching strategy is to switch the data stream from the primary path to a pre-calculated backup path in advance. The node that triggers the pre-switching strategy sends a path switching control message to the relevant nodes involved in the backup path. The path switching control message contains the complete node sequence or forwarding table entry information of the backup path. The node that receives the path switching control message pre-distributes the forwarding table entries of the backup path to its local data forwarding plane, so that when the data packet arrives, it can be forwarded immediately according to the forwarding table entries of the backup path, thereby achieving a seamless or fast switch from the primary path to the backup path.

[0044] Optionally, nodes maintain a dual forwarding table structure for the primary and backup routes, comprising two logical parts: a primary forwarding table and a backup forwarding table. When the primary path is functioning normally, packet forwarding is entirely based on the primary forwarding table, whose entries are generated and installed through routing protocols or path calculation processes. When a pre-switching policy is triggered, the entry in the backup forwarding table corresponding to the target backup path is activated, and the node switches its packet forwarding query from the primary forwarding table to the backup forwarding table. In some implementations, backup forwarding table entries can be dynamically calculated and installed when the pre-switching policy is triggered, or they can be pre-installed but inactive after path calculation is complete; triggering the pre-switching policy only requires modifying the priority or activation flag of the forwarding query. Pre-deploying backup forwarding table entries to the target node ensures rapid execution of the switching action, avoiding the delay caused by calculating a new path only after a path is interrupted.

[0045] In practice, the load balancing mechanism achieves this by adjusting forwarding priorities and defining a load balancing weight for each node. The calculation of the load balancing weight is based on the node's real-time resource status and remaining energy. Defined as the percentage of remaining energy of a node, and the queue length of a node. Defined as the ratio of the node's current data queue length to the node's maximum queue capacity. Load balancing weight. The percentage of remaining energy is calculated using a weighted average. and queue length percentage The specific formula is as follows: in the formula Representative node Load balancing weights, Representative node The percentage of remaining energy out of the total energy. Representative node The percentage of the total queue length to the length of the queue. and These are preset weighting coefficients used to balance the effects of energy and queue factors. Load balancing weights The higher the value, the stronger the node. The node is most suitable to undertake forwarding tasks at the current moment. The node periodically calculates its own load balancing weight. And broadcast it via signaling.

[0046] In some embodiments, when a node forwards data packets via the primary or backup path, if there are multiple selectable relay nodes for the next hop according to the forwarding table, the next hop is dynamically selected based on load balancing weights. The node collects the most recently broadcast load balancing weight values ​​of all selectable next-hop relay nodes and calculates the probability of each selectable next-hop relay node being selected based on these load balancing weight values. The probability of an selectable next-hop relay node being selected is proportional to its load balancing weight value; the specific probability allocation can employ weighted round-robin or random selection methods. For example, for a set of selectable next-hop relay nodes... Nodes in The probability of it being selected It can be calculated as follows: Based on a calculated probability distribution, nodes determine the next-hop node for each data packet using random number generation. This probability-based selection mechanism allows traffic to automatically skew towards nodes with higher load balancing weights, achieving dynamic traffic allocation based on real-time load conditions.

[0047] Example 5 In practice, network operation data collection is a continuous background process. Each node locally records key events and status indicators related to routing decisions and network performance. Nodes record link interruption events, including the identifiers of the two endpoints of the interrupted link, the precise timestamp of the interruption, the received signal strength indication just before the interruption, and the calibrated link interruption probability value. Nodes record path switching delay, defined as the time interval from when a node detects a primary path failure or triggers a pre-switching strategy until the node successfully forwards the next data packet via a backup path. Nodes record end-to-end packet loss rate, calculating the proportion of packets lost per unit time for a specific flow through transport layer acknowledgments or periodic probe messages. Nodes record energy consumption, periodically sampling their remaining energy and calculating the energy consumption rate per unit time. The recorded network operation data is stored in time-series format in the node's non-volatile memory, or periodically uploaded by the node to the management node in the network for centralized storage and analysis.

[0048] In some embodiments, the collected network runtime data is used to fine-tune the bidirectional long short-term memory network model online. The dataset used for fine-tuning is constructed from recorded data, where the input features of each training sample are a historical sequence of multidimensional state information, and the sample label is the corresponding actual link interruption state. This indicates that the interruption has not occurred. (Indicates an interruption). Online fine-tuning is typically performed on network management nodes or designated nodes with strong computing power. During fine-tuning, the current parameters of the pre-trained bidirectional long short-term memory (BSSM) network model are loaded, and a subset of the newly collected dataset is used as the fine-tuning training set. The parameters of the BSSM network model are updated slightly using the backpropagation algorithm and gradient descent optimizer, with the cross-entropy loss function as the optimization objective. After fine-tuning, the updated BSSM network model parameters are distributed to relevant nodes in the network, replacing the original model parameters, thus enabling the link interruption probability prediction capability to adapt to changes in the network environment.

[0049] It is understandable that adjustments to the link interruption probability threshold and control signaling cycle are based on long-term statistical results. Network management nodes or individual nodes themselves conduct long-term statistical analysis of recorded network operation data. These results include the success rate of path pre-switching strategies, average path handover latency, overall network energy consumption distribution, and the proportion of control signaling overhead. The link interruption probability threshold is adaptively adjusted based on these long-term statistical results. The adjustment logic is as follows: if statistics show that path handovers are too frequent but most abandoned old paths after handovers are not immediately interrupted, it indicates that the link interruption probability threshold is set too sensitively. In this case, the threshold should be appropriately increased to reduce unnecessary handovers. If statistics show that the number of path handovers is too low but a large number of unpredictable link interruptions occur, leading to packet loss, it indicates that the link interruption probability threshold is set too conservatively. In this case, the threshold should be appropriately decreased to trigger protective handovers earlier. The adjustment of the link interruption probability threshold follows a slow and gradual principle, with an upper limit on the magnitude of each adjustment.

[0050] Optionally, the update cycle of neighbor discovery messages and topology control messages can be adaptively adjusted based on long-term statistical results. Nodes monitor the frequency of network topology changes, which is quantified by the number of entries added or deleted from the neighbor table per unit time. Nodes monitor historical link stability data, which is evaluated by calculating the average or variance of the calibrated link outage probability of established links. Define an update cycle adjustment factor. Its calculation is related to the frequency of topology changes and the average link outage probability. A specific periodic adjustment formula is as follows: in, This represents the adjusted message update cycle. Represents the basic update cycle, adjustment factor It is a function of topology change frequency and average link stability. When the topology change frequency is high and the average link stability is poor, Take a negative value, such that Less than This shortens the update cycle and increases signaling frequency to quickly track changes. When the topology change frequency is low and the average link stability is good, Take a positive value such that Greater than This extends the update cycle and reduces signaling overhead. Through this dynamic adjustment, control response speed is improved during periods of network instability, while unnecessary energy and bandwidth consumption is reduced during periods of network stability.

[0051] In practice, the adaptive adjustment process of model parameters and operating thresholds can be triggered periodically or driven by specific events. Periodic triggering means that the system automatically executes a data aggregation, statistical analysis, model fine-tuning, and parameter update process every fixed long period (e.g., several hours or a day). Event-driven means that when network performance indicators (such as average packet loss rate, average end-to-end latency) continue to deteriorate and exceed a certain warning threshold, an adjustment process is immediately triggered. When fine-tuning the bidirectional long short-term memory network model online, it is important to retain a portion of the latest historical data as a validation set to evaluate the performance of the fine-tuned model and prevent overfitting. When distributing new bidirectional long short-term memory network model parameters or new threshold parameters, a version number management mechanism is used to ensure that the configuration used by each node in the network is consistent, and to allow for rollback to the previous stable version if problems occur during the update.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-altitude intelligent network routing optimization method based on link stability and energy sensing, characterized in that, include: Multidimensional status information is collected by nodes in the low-altitude intelligent network. The multidimensional status information includes node identifier, timestamp, three-dimensional coordinate position, remaining energy percentage, received signal strength indication, number of neighboring nodes, wind speed, heading angle, node status, packet loss rate, and link quality. Based on the multidimensional state information, a bidirectional long short-term memory network model is used to predict the link interruption probability, and the predicted value is subjected to probability calibration and smoothing to generate a calibrated link interruption probability. The link survival probability is calculated based on the calibrated link outage probability, and an outage risk adjustment factor and a comprehensive path weight model are defined by combining the node's remaining energy and link quality. The score of each candidate node is calculated based on the interruption risk adjustment factor, node coverage, and remaining energy. Multi-point relay nodes are elected according to the scores, and adaptive adjustment and coverage detection are achieved. The path cost is calculated based on the comprehensive path weight model, and the shortest path algorithm is used to select the primary path and backup path. When the calibrated link interruption probability exceeds a preset threshold, a pre-switching strategy is triggered. Maintain dual forwarding tables for primary and backup routes to enable data forwarding and load balancing based on nodes; The update cycle of neighbor discovery messages and topology control messages is dynamically adjusted based on the calibrated link outage probability.

2. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 1, characterized in that, The collection of multi-dimensional status information through nodes in the low-altitude intelligent network includes: Nodes periodically broadcast neighbor discovery messages, which include a unique node identifier, current timestamp, three-dimensional coordinates, remaining energy percentage, and a communication capability description. The communication capability description includes at least the supported channel bandwidth and modulation and coding scheme. After receiving the neighbor discovery message, other nodes update their local neighbor table and calculate the link packet loss rate and received signal strength indication. The receiving node integrates the wind speed, heading angle, and node status information it senses to construct a neighbor table containing all multi-dimensional status information.

3. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 2, characterized in that, The method of predicting link interruption probability using a bidirectional long short-term memory network model based on multidimensional state information includes: The multidimensional state information is scaled to generate an input vector that includes velocity magnitude, acceleration magnitude, angle vector, coordinates, energy percentage, link quality index, neighbor density, wind speed, and time interval. The input vector is fed into a pre-trained bidirectional long short-term memory network model, which is pre-trained using historical low-altitude network operation data. The original link interruption probability prediction value is output through forward and backward time-series processing. The original link interruption probability prediction value is probabilistically calibrated using the Platt scaling method and smoothed using an exponential moving average. The smoothing factor α of the exponential moving average is a configurable parameter with a value range of (0,1) and a default value of 0.

2.

4. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 3, characterized in that, The calculation of link survival probability based on the calibrated link outage probability includes: The link survival probability is defined as 1 minus the calibrated link failure probability; Calculate the average survival probability of adjacent edges, and combine a small constant to avoid division by zero error. Define the interruption risk adjustment factor as the link survival probability divided by the sum of the average survival probability and the constant. A comprehensive path weight model is constructed by combining link quality indicators and node remaining energy. The comprehensive path weight model is constructed using the integrated link quality indicators and the remaining energy of nodes, including: Define a link quality penalty, which is 1 minus the link quality metric, such that the worse the link quality, the larger the metric value. Define a penalty for the remaining energy of a node, which is 1 minus the percentage of the node's remaining energy, such that the lower the node's energy, the larger this value is; The link cost is defined as the linear sum of a risk cost, an energy penalty, and a quality penalty. The risk cost is the interruption risk adjustment factor multiplied by a first weighting coefficient. The energy penalty is the average of the remaining energy penalties of the nodes at both ends of the link multiplied by a second weighting coefficient. The quality penalty is the link quality penalty multiplied by a third weighting coefficient. The first, second, and third weighting coefficients are real numbers greater than zero. Their specific values ​​are preset through offline simulation or based on prior knowledge of network deployment, and can be slowly and adaptively adjusted based on the overall network status. Define the path cost of any candidate path, where the path cost is the sum of the link costs of all links on that path; The reciprocal of the path cost is defined as the path comprehensive weight. The higher the path comprehensive weight, the better the path.

5. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 4, characterized in that, The calculation of the score for each candidate node based on the interruption risk adjustment factor, node coverage, and remaining energy includes: The number of two-hop neighbors covered by a candidate node is used as a coverage metric. Obtain the remaining energy of the candidate nodes; Calculate the average outage risk adjustment factor between the candidate node and its neighboring nodes; Candidate node scores are obtained by weighted summation of coverage, remaining energy, and average interruption risk adjustment factor.

6. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 5, characterized in that, The process of electing multi-point relay nodes based on scores and implementing adaptive adjustment and coverage detection includes: Select the candidate node with the highest score as the multi-point relay node; The coverage of two-hop neighbors is checked periodically. The period of the periodic check can be dynamically set according to the average network mobility or topology change rate. When the coverage is insufficient, the relay node set is dynamically adjusted. Verify relay node paths to avoid loop formation; The selection of the candidate node with the highest score as the multi-point relay node includes: In the set of one-hop neighbors of the current node, sort them according to the scores calculated by the candidate node scoring function; Select the node with the highest score from the sorting results as the preferred multi-point relay node for this node; The identifier and score of the preferred multipoint relay node are broadcast to neighbors through relay node election messages. If a node receives election messages from multiple neighbors all electing it as the preferred relay node, then the node confirms itself as a multipoint relay node and broadcasts a relay announcement message to the entire network. If a node receives multiple relay announcement messages, it determines its own assigned relay node based on the rating of the announcement nodes, thus completing the establishment of a multi-point relay topology.

7. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 6, characterized in that, The calculation of path cost based on the comprehensive path weight model includes: The risk cost term is defined based on the calibrated link outage probability. Define energy penalty terms based on node remaining energy penalties; Define quality penalty items based on link quality penalties; The link cost is obtained by linearly adding the risk cost, energy penalty, and quality penalty, and the path cost is the sum of all link costs.

8. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 7, characterized in that, The method of using the shortest path algorithm to select the primary path and backup path includes: Apply Dijkstra's algorithm to find the main path with the minimum path cost; At the same time, several backup paths are built as backups; When the probability of link interruption after calibration of the primary path exceeds the threshold, the pre-switching strategy is triggered to switch to the backup path; The application of Dijkstra's algorithm to find the main path with the minimum path cost includes: Starting from the source node, initialize its path cost to zero, and set the initial path cost of all other nodes in the network to infinity. Create a set of nodes for which the shortest path has not yet been determined, initially containing all nodes in the network; In each iteration, the node with the lowest known path cost is selected from the set as the current processing node; Iterate through all neighboring nodes of the current processing node. For each neighboring node, calculate the total path cost from the source node to that neighboring node via the current processing node. This total path cost is equal to the known path cost of the current processing node plus the link cost connecting the two nodes. If the calculated total path cost is less than the known path cost of the neighboring node, then update the path cost of the neighboring node to the total path cost, and record the currently processed node as its predecessor node. Remove the currently processed node from the set of nodes with undetermined shortest paths; Repeat the above iterative process until the target node is removed from the set. At this point, based on the predecessor node information recorded by each node, backtrack from the target node to the source node to obtain the main path with the minimum path cost.

9. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 8, characterized in that, The maintenance of dual forwarding tables for primary and backup routes includes: When the main router is functioning normally, data is forwarded through the main forwarding table. When the pre-swapping policy is triggered, the forwarding table entries of the backup path are sent to the target node in advance. The advance sending operation is executed immediately after the pre-swapping policy is triggered. Load balancing is achieved by adjusting forwarding priorities based on node remaining energy and queue length. Specifically: The method of adjusting forwarding priority based on node remaining energy and queue length to achieve load balancing includes: A load balancing weight is defined for each node. The load balancing weight is calculated by weighting the node's remaining energy percentage and queue length, where the queue length percentage is the ratio of the current data queue length to the maximum queue capacity. When a node forwards a data packet through the primary path or backup path, if there are multiple relay nodes available for the next hop, the node is selected based on the load balancing weight value currently broadcast by each relay node, with the node having a higher weight value having a greater probability of being selected. Nodes periodically calculate and broadcast their latest load balancing weights to achieve dynamic traffic distribution based on real-time load status.

10. The low-altitude intelligent network routing optimization method based on link stability and energy awareness according to claim 9, characterized in that, The method further includes: Collect network operation data, and update the parameters of the bidirectional long short-term memory network model and adjust the preset threshold accordingly; specifically including: Record link interruption events, handover latency, packet loss rate, and energy consumption data; The bidirectional long short-term memory network model is fine-tuned online using recorded data. The online fine-tuning is triggered periodically or when the model's prediction error continues to exceed a set threshold. The link interruption probability threshold and control signaling cycle threshold are adaptively adjusted based on long-term statistical results, including path pre-switching success rate, average handover delay, and control signaling overhead ratio.

Images