A Joint Optimization Method for FSO / RF-UANET Routing and Localization Based on Link State Aware Spatiotemporal Coordination

By introducing a link-state-aware spatiotemporal collaborative FSO/RF-UANET routing and localization joint optimization method into mobile ad hoc networks, the shortcomings of traditional anti-interference methods in handling heterogeneous link attribute differences are solved, and efficient communication and stable transmission of the network are achieved in complex environments.

CN123124047APending Publication Date: 2026-11-17NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611297345.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to balance environmental adaptability, real-time decision-making, and system-level verifiability in mobile ad hoc networks. In particular, under dynamic spectrum contention and intelligent reactive interference, traditional anti-interference methods cannot effectively handle the differences in heterogeneous link attributes, leading to low network energy utilization and difficulties in clock synchronization.

Method used

A joint optimization method for FSO/RF-UANET routing and localization based on link state awareness and spatiotemporal coordination is proposed. By constructing a joint resource scheduling model with discrete periodic interaction, frequency switching overhead and jammer detection and tracking mechanism, and adopting the attention-enhanced PER-D3RQN algorithm, the method enhances the ability to capture key historical adversarial information and make decisions in complex action spaces, thereby achieving joint optimization of channel, transmit power and rate level.

Benefits of technology

In an intelligent reactive interference environment, it improves cumulative rewards and throughput, enhances link transmission efficiency, and maintains good network stability and communication quality in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124047A_ABST
    Figure CN123124047A_ABST
Patent Text Reader

Abstract

This invention discloses a joint optimization method for FSO / RF-UANET routing and localization based on link state-aware spatiotemporal coordination. This method constructs a joint resource scheduling model including discrete periodic interactions, frequency switching overhead, and jammer detection and tracking mechanisms. It proposes an Attention-enhanced PER-D3RQN (APER-D3RQN) anti-jamming method, extending anti-jamming decision-making from a single channel selection to joint optimization of channel, transmit power, and rate level, enhancing the ability to capture key historical adversarial information and make decisions in complex action spaces. Simulation results show that APER-D3RQN achieves higher cumulative rewards and throughput in intelligent reactive jamming environments, and exhibits superior performance in metrics such as actual jamming hit rate, tracking trigger ratio, and channel switching rate, maintaining good link transmission efficiency while reducing the risk of jamming exposure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination, belonging to the field of computer communication technology. Background Technology

[0002] Mobile Ad Hoc Networks (MANETs) have significant application value in emergency communication, unmanned system collaboration, and complex adversarial scenarios. However, their open and shared wireless propagation medium is susceptible to external interference, especially under dynamic spectrum competition and intelligent reactive jamming conditions. Traditional anti-jamming methods relying on prior rules or static parameter configuration struggle to balance environmental adaptability, real-time decision-making, and system-level verifiability. Deep Reinforcement Learning (DRL) provides a new approach to anti-jamming for ad hoc networks in complex electromagnetic environments by achieving online policy optimization through interaction with the environment. However, the common challenges in time-varying and reactive jamming communication adversarial processes, such as observable state components, strong temporal correlation, and high dimensionality of action decisions, still necessitate the development of system models that more closely resemble real communication processes and more effective decision-making algorithms. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing routing protocols, such as lack of adaptability in handling heterogeneous link attribute differences, low network energy utilization, and difficulty in clock synchronization under high dynamic environments. It proposes a joint optimization method for FSO / RF-UANET routing location based on link state-aware spatiotemporal coordination. This method constructs a joint resource scheduling model incorporating discrete periodic interactions, frequency switching overhead, and jammer detection and tracking mechanisms. It also proposes an Attention-enhanced PER-D3RQN (APER-D3RQN) anti-jamming method, expanding anti-jamming decision-making from a single channel selection to joint optimization of channel, transmit power, and rate level, enhancing the ability to capture key historical adversarial information and make decisions in complex action spaces. Simulation results show that APER-D3RQN achieves higher cumulative rewards and throughput in intelligent reactive jamming environments, and exhibits superior performance in metrics such as actual jamming hit rate, tracking trigger ratio, and channel switching rate, maintaining good link transmission efficiency while reducing the risk of jamming exposure.

[0004] The technical solution adopted by this invention to solve its technical problem is: a joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination, which includes the following steps: Step 1: Physical modeling of node features; In FSO / RF-UANET, it is necessary to consider the high dynamic characteristics, energy-constrained characteristics, and heterogeneity of UAV nodes, and to transform these physical conditions into multi-dimensional feature attributes of graph nodes. Any node in the network... In the time slot The dynamic physical characteristic states are modeled as high-dimensional feature tensors. ; Step 1-1: Three-dimensional spatial position representation : ; Step 1-2: Maneuvering Inertial Characteristics : The instantaneous velocity of the drone node in each coordinate axis direction, ; Steps 1-3: Node Normalized Residual Energy Characteristics : This refers to the remaining energy at each node, using the initial energy. Normalize it. ; Steps 1-4: Node Role Identifiers One-Hot encoding is used to distinguish the roles of nodes (ground signal stations and drone nodes). Step 2: Physical modeling of edge features; When constructing the input to a graph neural network, not only are node feature matrices defined, but edge feature tensors are also constructed, with each edge... eigenvectors It includes the following key physical quantities; Step 2-1: Geometric distance features : This is the fundamental variable that determines the path loss in free space; Step 2-2: Jitter Variance Prediction : ; Steps 2-3: Historical Channel States Used to record the past The signal-to-noise ratio of each time slot serves as a key characteristic for verifying link reliability. Steps 2-4: Link Mode Identification Used to distinguish between FSO links and RF links; Steps 2-5: Link Scenario Identification One-Hot coding is used to distinguish link types, including air-to-air links and air-to-ground links, because different types of links have different divergence angles and receiver aperture parameters. Step 3: Link state-aware attention mechanism; HA-GNN employs a node-centric attention mechanism, and its attention coefficient... Based solely on node features The decision was made, but this attention mechanism is not suitable for FSO / RF-UANET networks. The E-MA-HGNN algorithm introduces an edge fusion layer, which enables the attention mechanism to focus on the heterogeneous properties of the links. First, the node and edge features are projected using a learnable linear transformation matrix: (4.40) (4.41) in, Indicates different link types, In the E-MA-HGNN sub-algorithm, the source node, target node, and the edge features between them are concatenated into ternary pairs to construct the attention coefficients. : (4.42) in, This indicates a feature concatenation operation. Indicates the first One's attention, when When dealing with a single FSO link and high jitter, the E-MA-HGNN sub-algorithm will learn through training and assign a larger negative weight to this feature, thereby reducing jitter. This will allow the router to avoid selecting high-bandwidth but extremely unstable FSO links, thus intelligently switching to RF links, and when Corresponding to one RF link and distance When the path loss is large, it becomes a critical factor. The E-MA-HGNN sub-algorithm reduces its weight, and routing will avoid selecting long-distance, low-speed RF links. For different networks, network quality is measured in multiple dimensions; sometimes low latency is required, and sometimes high jitter is needed. To capture these characteristics simultaneously, a method is adopted... Each attention head has its own independent attention head, and different attention heads focus on different features. The first attention head may focus on distance features, while the second attention head may focus on distance features. An individual's attention may focus on jitter features; Step 4: Policy gradient training; Define the objective function of the E-MA-HGNN sub-algorithm as the policy. Expected value of the expected reward: (4.46) in, This represents a complete relay path. This is the reward corresponding to that path; Since routing decisions are discrete actions, it's impossible to directly calculate the mathematical derivative for selecting actions at discrete next-hop nodes. Therefore, the standard logarithmic derivative is used to generate the policy gradient, and the agent is in the state... Below, the decoder outputs the probability of selecting a specific action. Recorded as ; Used to optimize the objective function parameters The policy gradient is described as follows: (4.47) in, This represents the total number of hops for this relay path. This refers to the state of the decoder network. Select node for output The probability of; Therefore, E-MA-GNN The parameters can be updated as follows: (4.48) in, The learning rate is the value of the E-MA-GNN sub-algorithm. After using the E-MA-HGNN sub-algorithm to solve the sub-problem of how to select the optimal relay path, the EST-UL-GNN sub-algorithm is used to solve the UAV positioning and synchronization optimization sub-problem. Step 5: Elastic spatiotemporal coupling; To prevent FSO link interruption, the EST-UL-GNN sub-algorithm proposes to treat the UAV's spatial position and clock synchronization state as a pair of strongly coupled variables. By sacrificing some spatial degrees of freedom, the UAV node is moved to a more advantageous position to achieve stability of time synchronization. It can also maximize spatial coverage and increase spatial degrees of freedom when the clock is stable. The method defines the system state as a composite vector. ; spatial state The spatial state of the UAV node is a key factor determining the geometric gain and propagation delay of the FSO link. The first... The spatial state of each node is: (4.49) in, These represent the position, instantaneous velocity, and acceleration of the UAV node in a three-dimensional Cartesian coordinate system, respectively. Step 6: Dual-stream LSTM network architecture; The EST-UL-GNN algorithm employs a learning architecture based on a dual-stream structure, which includes a spatial stream and a temporal stream, with the two interacting deeply on features. Spatial flow is used to model the physical motion and network topology of UAVs. Its input mainly consists of multi-dimensional time-series inputs, including historical trajectory sequences. And global topological embeddings extracted by the preorder graph neural network Unlike traditional unidirectional LSTM, Spatial Flow uses a bidirectional long short-term memory network (Bidirectional LSTM) as its core encoder, which can simultaneously capture the inertial laws of UAV motion and the constraints of target position. It can use past positions and combine them with neighbor position information extracted by graph neural networks to predict the ideal position that maximizes network connectivity. The state update formula is: (4.54) in, yes The set of all internal weight matrices, eigenvectors This indicates the spatial requirements for drone nodes to maintain optimal link connectivity in the future. The time stream is used to handle the hidden variables of physical layer clock evolution, mainly the physical characteristics of the crystal oscillator. Its inputs include past clock offsets. With key environmental parameters, the input vector is Because clock drift has strong unidirectional causality, the time flow will be modeled using a unidirectional LSTM: (4.55) in, It is the set of weight parameters for a unidirectional LSTM. Step 7: Design a multi-objective unsupervised loss function; The loss function transforms the physical laws of link communication into optimization constraints. The EST-UL-GNN algorithm, through self-deduction, explores the spatiotemporal equilibrium point, and the loss function... Defined as a weighted sum across four dimensions: (4.57) in, To adjust the parameters of each weight, This is used to minimize the sum of squared distances from all ground signal stations to the drone node, ensuring basic geometric coverage: (4.58) in, It is a ground signal station coordinates It is the position adjustment amount output by the model. This allows drone swarms to cover areas with dense ground signal stations. Link quality loss Link quality loss is used to maximize the average signal-to-noise ratio of the entire network. It utilizes the edge features calculated by the E-MA-HGNN sub-algorithm to maximize the average link score of the entire network. (4.59) To prevent time slot misalignment caused by nonlinear drift of the crystal oscillator, an exponential penalty term is introduced: (4.60) in, This represents the clock skewness of future moments predicted by the time-flow LSTM. This indicates the maximum allowable gate deviation threshold of the system. It is a decay coefficient used to adjust the sensitivity to the risk of synchronization loss.

[0005] Beneficial effects: 1. This invention proposes a joint optimization algorithm for FSO / RF-UANET routing and localization based on link state awareness and spatiotemporal coordination. The E-MA-HGNN sub-algorithm constructs node features and edge attribute features, and introduces RF / FSO link modes, channel quality, jitter variance and node remaining energy into the graph attention calculation process to perceive the attribute differences of different links, so that routing decisions can shift from passive rule switching to active mode selection based on link state.

[0006] 2. The EST-UL-GNN sub-algorithm of this invention considers the coupling relationship between UAV position changes and clock drift. It uses dual-stream LSTM to output position adjustment results through an elastic coupling mechanism, and the two achieve joint optimization. Through the collaboration and interaction of the E-MA-HGNN and EST-UL-GNN dual models, it can dynamically balance the stability of the RF link and the high bandwidth of the FSO link, and maintain high communication quality in complex environments.

[0007] 3. Through simulation experiments, it is verified that in FSO / RF-UANET, each sub-algorithm of the E-MA-HGNN & EST-UL-GNN joint optimization algorithm plays an important role, and the two work together to show superior comprehensive performance on multiple performance indicators. Attached Figure Description

[0008] Figure 1 This invention provides a deployment scenario diagram for FSO / RF-UANET.

[0009] Figure 2 This invention presents an FSO / RF-UANET architecture diagram.

[0010] Figure 3The present invention presents a framework diagram of a joint optimization algorithm for FSO / RF-UANET routing and localization based on link state awareness and spatiotemporal coordination.

[0011] Figure 4 This invention provides end-to-end delay diagrams for different node sizes.

[0012] Figure 5 This invention relates to packet delivery rates under different node sizes.

[0013] Figure 6 This invention relates to network throughput under different node sizes.

[0014] Figure 7 This invention relates to end-to-end energy consumption at different node scales.

[0015] Figure 8 Comparison chart of mode switching quantities under N60-M10 network scale of this invention.

[0016] Figure 9 This invention compares the resilience and recovery capabilities of networks with scales of N60 to M10.

[0017] Figure 10 This invention provides a comprehensive evaluation diagram of link status under different node sizes. Detailed Implementation

[0018] The invention will now be described in further detail with reference to the accompanying drawings.

[0019] The framework of this invention mainly comprises three core mechanisms: First, at the routing decision level, the algorithm explicitly models node and edge features and introduces an edge-enhanced graph attention mechanism to distinguish link modes and link quality, assigning different forwarding weights to different links. This transforms passive hard handover into active mode selection based on deep features. Second, at the positioning and synchronization level, the algorithm introduces a flexible spatiotemporal optimization mechanism. By constructing a dual-stream LSTM network, it achieves flexible optimization of UAV position and synchronization strategies, supporting stable communication and efficient task collaboration for UAVs in complex and ever-changing environments. Finally, through sequential training and collaborative inference of the dual models, the algorithm exhibits superior overall performance.

[0020] In sudden disasters, drone swarms often need to transmit massive amounts of high-definition video streams, which laser ad hoc networks can easily handle. However, disaster sites such as fires and floods are often accompanied by dense smoke and thick clouds. These complex suspended particles severely block the penetration of laser signals, causing a precipitous drop in the signal-to-noise ratio of the laser link, and even leading to the complete paralysis of the entire laser ad hoc network. A heterogeneous drone ad hoc network with a dual-mode communication link can sense changes in the surrounding weather. When encountering dense smoke obstruction, it will decisively switch critical command data to the radio frequency link for transmission. This intelligent switching mechanism ensures uninterrupted communication between the rescue command center and the disaster area frontline.

[0021] Based on the aforementioned application scenarios, this invention constructs a heterogeneous UAV ad hoc network (FSO / RF-UANET) system model incorporating both free-space optical (FSO) and radio frequency (RF) dual-mode communication links. The network consists primarily of UAVs and ground signal stations, equipped with FSO transceivers and RF communication modules. In this network topology, each node can communicate with other nodes using the FSO link, thereby collaboratively achieving multi-hop transmission of network data. When the signal-to-noise ratio (SNR) falls below a certain threshold, a switching mechanism is employed to switch to the RF link. RF communication offers advantages in handling adverse weather conditions and non-line-of-sight transmission, enabling the network to maintain stable operation even under obstructed line-of-sight or complex environments, ensuring continuous network communication. Furthermore, each UAV node model integrates a navigation and timing module, a clock-keeping module, and an environmental perception module, thus providing the UAV with high-precision flexible clock synchronization in addition to supporting basic FSO / RF communication payloads.

[0022] Based on the aforementioned heterogeneous self-organizing network, adding an energy state to the drone can simplify it into... Figure 2 .according to Figure 2 The highly dynamic heterogeneous UAV self-organizing network is modeled as a time-varying weighted undirected graph. The set of nodes in the graph Represents ground signal station and drone nodes A set, each node Each drone node is equipped with an FSO transceiver and an RF omnidirectional antenna, enabling it to process two modes of signals simultaneously. Each drone node in the network possesses a certain amount of energy. This is used for laser communication and route maintenance. Furthermore, the node's position changes continuously in three-dimensional space, denoted as... Unlike traditional unimodal networks, the graph contains a heterogeneous set of edges. It consists of RF subsets and FSO subsets. During forward propagation, the RF link offers omnidirectional coverage and strong anti-interference capabilities, but its bandwidth is limited and it is susceptible to multipath interference. When the distance between nodes... A link is considered to exist if its radius is less than the maximum communication radius of the RF link. FSO links have extremely high bandwidth but are highly sensitive to weather and jitter, and only function when the following conditions are met. Furthermore, the link exists when there is no obstruction at the line of sight and the reception error is less than the receiving field of view. During backpropagation, the discrete topology generation process is transformed into a continuously differentiable graph reconstruction process. For any node in the network... Its reconstructed differentiable adjacency matrix elements Expressed as (4.1) in, This represents the smoothing temperature hyperparameter.

[0023] Considering the characteristics of networks, for any node pair The links are all bidirectional, meaning the nodes... and nodes They can send laser signals to each other.

[0024] This invention describes RF channel modeling in detail, and models air-to-ground channel model and air-to-air channel model respectively.

[0025] Unmanned Aerial Vehicle (UAV) Air-to-Ground (A2G) Channel Model The A2G channel serves as the transmission link between UAV nodes and ground signal stations. Its core challenge lies in quantifying the nonlinear relationship between elevation angle and path loss. To accurately model the A2G communication environment, this invention constructs a comprehensive channel model and elaborates on its key parameters, including the probabilistic line-of-sight model, path loss exponent, shadowing fading, and small-scale fading distribution.

[0026] Probabilistic line-of-sight model In existing A2G channel research, this invention will adopt the probabilistic line-of-sight model proposed by Al-Hourani et al. This model, based on statistical geometry principles, models the average path loss of an A2G link as a weighted sum of Loss (LoS) path loss and Non-Line-of-Sight (NLoS) path loss.

[83] .

[0027] The probability of line-of-sight existence on LoS links It is the angle of elevation. The sigmoid function is expressed as: (4.2) in, and These are empirical parameters determined by the specific propagation environment, and are usually related to environmental types such as cities, suburbs, and densely populated cities.

[0028] When elevation angle When the drone is small, it is at low altitude or long distance, and the signal must penetrate multiple layers of buildings. When the elevation angle approaches 0, the channel is dominated by NLoS. In addition, the drone was positioned directly above the ground signal station, avoiding view of most buildings. Approaching 1.

[0029] Based on the above Average path loss It can be represented as: (4.3) (4.4) (4.5) in, It is free space path loss. and These represent the average additional losses under Loss and Non-LoS conditions, respectively. Typically, in densely populated cities... Much larger This reflects the effects of penetration loss and diffraction loss.

[0030] Path loss exponent and shadow fading As the drone's altitude increases, the impact of ground scattering gradually weakens, and the path loss index decreases. It gradually decreases, approximating the free-space value from a typical ground value. The path loss exponent can be modeled as: (4.6) As shown in the above formula, increasing the drone's altitude reduces the path loss exponent, but its drawback is the increased physical distance. Therefore, a system design trade-off is needed to find an optimal altitude that minimizes the total path loss within a specific coverage radius.

[0031] Shadow fading describes the slow fluctuations of a signal when it encounters a large obstacle, and follows a log-normal distribution. In an A2G channel, the standard deviation of shadow fading... It exhibits a strong angle dependence. The angles are divided into low elevation angles and high elevation angles. Specifically, at a low elevation angle, the signal will pass through a large number of buildings, and the shadows will produce violent fluctuations. The effect is relatively large. However, at high elevation angles, the signal path is mostly a line-of-sight path, and the severe shadowing effect is no longer produced. Significantly reduced. This invention employs an exponential decay model to describe this phenomenon.

[84] : (4.7) Small-scale fading Small-scale fading describes the rapid fluctuations of a signal over a short distance. In terrestrial communications, due to the complexity of multipath propagation and the lack of strong line-of-sight (LoS) components, Rayleigh fading is typically used to describe it. However, in A2G channels, due to the altitude advantage of UAV nodes, strong LoS components exist. Therefore, this invention uses a Rician fading model to describe the channel characteristics.

[0032] The core parameter of the Rice distribution is The factor is defined as the ratio of Loss-of-Sight (LoS) power to the power of the multipath scattering components. A significant characteristic of A2G channels is... The factor is not described by a constant, but rather by the ground elevation angle of the UAV node. Functions: (4.8) In open environments with few distractions, regardless of altitude... The factor remains consistently high and stable. However, in complex environments, The factor changes dramatically with the angle of elevation: when the angle of elevation... Approaching hour, At this point, multipath components can be ignored, and the channel approaches an additive Gaussian channel (AWGN); when the elevation angle When it approaches 0, This results in complex multipath propagation and the absence of strong line-of-sight (LoS) components, causing the channel to degrade into Rayleigh fading.

[85] .

[0033] Furthermore, even though ground signal stations move at low speeds or remain stationary, the high-speed flight of drones introduces a significant Doppler shift. The maximum Doppler shift is: (4.9) It directly affects the coherence time of the channel. For the Sub-6GHz band, this is typically in the range of several hundred hertz, but in the millimeter-wave band, the Doppler shift can reach several thousand hertz, requiring the communication system to have extremely fast channel estimation and equalization capabilities.

[0034] Unmanned aerial vehicle (UAV) air-to-air (A2A) channel model Air-to-air (A2A) channels are primarily used for communication between UAV nodes. Compared to A2G channels, A2A channels have no surrounding obstacles and very few background scatterers, but interference from far-field reflections from the ground and scattering by the UAV itself cannot be ignored.

[0035] transmission mechanism Due to the lack of scattering from surrounding buildings and the exclusion of complex statistical fading models, the A2A channel can be modeled as a two-ray model or a multi-ray model.

[0036] The most basic A2A model considers the direct path (LoS) and a ground reflection path. Received signal power. The interference exhibits alternating wavy characteristics of constructive and destructive interference as distance changes, depending on the phase difference between the reflected and direct waves.

[86] .

[0037] When the ground is relatively flat, specular reflection is significant. The model needs to consider the ground's dielectric constant and reflection coefficient. In more complex scenarios, the model is extended to a multi-ray model, incorporating the Loss of Space (LoS), ground specular reflection, and diffuse scattering components from surrounding drones. This multi-ray model can more accurately capture fluctuations in signal strength. (Received signal) Represented as: (4.10) in, For the range of Loss, For the first The amplitude of the reflection path, It includes phase changes caused by high-speed motion.

[0038] Statistical fading model This invention requires a rapid and accurate assessment of link quality under large-scale dynamic topologies. While multi-ray models can accurately describe the physical propagation mechanism under specific trajectories, their high computational complexity makes them unsuitable for network layer training and derivation. This invention employs the statistical model Nakagami-m distribution to model small-scale fading in A2A links due to its highly flexible fitting properties, achieved through parameter... Control the shape of the distribution.

[0039] (4.11) A2A channel The value is typically positively correlated with the distance and altitude between drones. As altitude increases, multipath propagation decreases. As the value increases, the channel quality improves.

[0040] Relative Doppler frequency shift In an A2A channel link where both ends are moving drones, the relative velocity between the two ends causes a Doppler frequency shift. The amount of this frequency shift... The calculation is as follows: (4.12) in, It is a velocity vector. It is a directional unit vector.

[87] .

[0041] Communication energy consumption model In FSO / RF-UANET, the energy of the UAV is a key factor in network lifetime and communication link stability. This invention establishes an accurate communication energy consumption model to precisely describe the energy consumption of each UAV node. (UAV node) At any moment Instantaneous total power consumption It mainly consists of two parts: the power used to maintain the drone's flight. and communication power used for data transmission and link maintenance. .

[0042] (4.13) in, This represents the basic static power consumption during the operation of the drone.

[0043] UAV propulsion energy consumption model For rotary-wing UAVs operating in FSO / RF-UANET, propulsion power consumption is typically much greater than communication power consumption, and its relationship with flight status is non-linear. This invention employs a closed aerodynamic power model, which describes propulsion power... With flight speed Relationship

[88] : (4.14) in, and These represent the blade drag power and induced power of the UAV in hovering state, respectively. Indicates the tip velocity of the rotor blades. This represents the average induced velocity during hovering. The drag coefficient of the fuselage. air density, For rotor realism, The rotor disk area, This refers to the instantaneous flight speed of the drone.

[0044] Heterogeneous network communication power consumption model In FSO / RF-UANET, communication power consumption varies significantly across different modes. (Communication power) It depends not only on the data transmission rate, but also on the physical characteristics of the current mode link, such as the power amplifier in the RF link and the alignment in the FSO link. The communication power model is expressed as... (4.15) in, This indicates the power consumption of the basic circuitry of the communication module. Represents a node The set of neighboring nodes, Let these represent the existence variables of the RF and FSO links, respectively, satisfying mutual exclusion constraints. This means that a link can only use one mode at a time.

[0045] RF link power consumption The power consumption of an RF link is primarily concentrated in the power amplifier (PA), and is strongly correlated with distance. This power consumption is mitigated by path loss and maintaining the target signal-to-noise ratio at the receiver. Required RF radiated emission power for (4.16) in, For noise power spectral density, For bandwidth, This is the path loss index (typically 2 to 2.5 for A2A channels). These are the transmit and receive antenna gains, respectively. Power amplifier efficiency is taken into consideration. In addition to the fixed losses of the RF circuit, the total circuit power consumption model of the RF link is as follows: (4.17) This model shows that the energy efficiency of RF links drops sharply during long-distance transmission.

[0046] FSO link power consumption Unlike RF links, the energy consumption bottleneck of FSO links lies not in laser emission power, but in the ATP system. To maintain precise narrow-beam alignment during UAV flight, the servo motors and fast mirrors (FSM) of the ATP system continuously consume significant mechanical communication energy. The FSO link power consumption model is approximately expressed as follows: (4.18) in, This represents the laser emission power, which is usually a constant. This represents the dynamic power consumption during the capture, alignment, and tracking process of the ATP system.

[89] .

[0047] According to the above formula, at short distances or low speeds, due to While FSO links are more energy-intensive than RF links, they are significantly more energy-efficient than RF links for ultra-high-speed or long-distance transmission.

[0048] Node energy state update node Remaining energy This is a key state variable in the E-MA-HGNN routing decision. Since the power consumption of UAV propulsion is much greater than that of link communication, if propulsion energy consumption is included in the end-to-end link evaluation, the impact of route selection and FSO / RF mode selection on communication energy consumption is easily masked by flight energy consumption. Therefore, the remaining energy does not consider UAV propulsion energy consumption, and the energy evolution over time follows an integral form. (4.19) In discrete-time simulation, the first The remaining energy in each time slot is updated to (4.20) To avoid capacity exhaustion in FSO / RF-UANET, when the normalized remaining energy of a node... Below the set threshold In this case, the E-MA-HGNN algorithm will actively reduce the probability of a node being selected as a relay by using the energy weight term in the edge features, thereby extending the overall lifespan of the network.

[0049] Problem Modeling and Analysis Problem Modeling In the high-dynamic networking of unmanned aerial vehicles (UAVs), the network not only needs high throughput but also comprehensive performance such as timeliness, reliability, stability, and durability. Therefore, this invention considers six key indicators to measure routing performance: average end-to-end latency, packet delivery rate, mode switching volume, network throughput, end-to-end communication energy consumption, and resilience. The core optimization problem is: how to find an optimal path between a given source node and destination node in a heterogeneous topology that achieves the best solution in terms of timeliness, reliability, stability, and durability. A comprehensive evaluation function was proposed. The routing performance is determined by the following routing evaluation function, as shown in the equation: (4.21) in, For topological state tensors, This represents a feasible route from the source node to the destination node. The weights of each evaluation indicator are adaptively adjusted, and the weight coefficients also need to be dynamically adjusted in different scenarios.

[0050] End-to-end delay evaluation function End-to-end latency is a core physical quantity for measuring the real-time response capability of a network. In FSO / RF-UANET, path latency is... The total latency includes not only the signal propagation time and node processing time, but is also constrained by the physical characteristics of the link modes. According to the heterogeneous topology model definition in Section 4.2.1, the link has modal identifier variables. Single-hop physical delay It consists of transmission delay, propagation delay, and hardware warm-up alignment delay for a specific mode: (4.22) in, For data packet size, The physical bandwidth corresponding to the current link mode. Additional time delays specific to a mode (e.g., the ATP mechanical alignment time of FSO).

[0051] Combined with the average queuing delay of nodes To eliminate the influence of absolute values ​​on dimensions, the time delay evaluation function is defined based on the maximum tolerable time delay. Exponential decay form: (4.23) in, This is the latency penalty sensitivity coefficient. The closer the value of this function is to 1, the better the real-time transmission performance of the path.

[0052] Package delivery rate evaluation function The packet delivery rate is a direct indicator of the reliability of a link transmission. Links with different modes and location relationships follow different probability fading distributions: for air-to-ground (A2G) radio links, channel fading is determined by a probabilistic line-of-sight model. It is dominated by the Rician distribution; for air-to-air (A2A) radio frequency links, it follows the Nakagami-m fading distribution; while for FSO links, the intensity fluctuations are affected by the log-normal turbulence distribution caused by atmospheric convection and the pointing error.

[0053] Define the probability of a single-hop link succeeding as: It is when the signal-to-noise ratio (SNR) at the receiver is greater than the demodulation threshold. Probability integral: (4.24) in, Let be the probability density function.

[0054] The packet delivery rate evaluation function is defined as the product of the independent connectivity probabilities of each hop on the path: (4.25) Mode switching overhead evaluation function Mode switching overhead arises from the high-frequency, blind "ping-pong switching" between heterogeneous modes. The mode switching overhead evaluation function is defined as: (4.26) in, This represents the switching overhead penalty coefficient. When frequent unordered mode switching occurs in the links of the path, the difference accumulation term surges, causing the function's score to decrease.

[0055] Network throughput evaluation function According to Shannon's theorem and maximum flow theory, the theoretical effective throughput of an end-to-end path is limited by the bottleneck links on that path. The end-to-end bottleneck throughput capacity of the path is defined as: (4.27) To perform a normalized evaluation in multi-objective optimization, it is compared with the maximum throughput under ideal conditions. Perform normalization mapping: (4.28) This objective function drives the algorithm to proactively assign higher attention weights to FSO links during decision-making, and to improve the weakest hop by shortening the physical distance of a single hop. .

[0056] End-to-end communication energy consumption evaluation function FSO / RF-UANET lifetime is directly limited by the energy consumption of the UAV. Relay UAV nodes with sufficient remaining energy should be prioritized. The end-to-end energy consumption assessment function is defined as: (4.29) in, This represents the initial energy load of the node. As the algorithm successfully switches traffic from nodes that are about to run out of power to nodes with sufficient power, the value of this evaluation function will steadily increase, thereby achieving energy load balancing at the network level.

[0057] Damage resistance evaluation function The resilience performance demonstrates the topology self-healing capability of FSO / RF-UANET in the event of sudden node failures or interference-induced crashes.

[0058] Throughput retention rate is defined as: (4.30) in, The average throughput before the failure occurred. The first time after the failure occurred throughput at any given moment To avoid extremely small constants with a denominator of zero.

[0059] The network resilience evaluation function is defined as follows: (4.31) in, This indicates the throughput retention rate at the moment of failure. This indicates that the observation window has been restored. This represents the time required for throughput to recover to a set percentage. A larger value indicates a smaller throughput drop and faster recovery speed after a sudden node failure.

[0060] According to equation (4.21), the core optimization objective of this invention can be expressed as a comprehensive performance optimization of FSO / RF-UANET, and equation (4.32) defines the optimization objective function: (4.32) (4.33) (4.34) (4.35) (4.36) (4.37) (4.38) (4.39) in, These are adaptively adjusted weights for each evaluation indicator, satisfying the normalization condition. This covers the six core physical quantity evaluation sub-items mentioned above. Formula (4.33) constrains that the sum of all adaptive adjustment weight coefficients must be exactly equal to 1, and each independent weight coefficient cannot be less than 0. Formula (4.34) constrains that the data traffic transmitted on any link cannot exceed the maximum physical throughput of the current communication mode. Formula (4.35) constrains that the total inflow traffic received by the relay node must be equal to the total outflow traffic forwarded. Formula (4.36) constrains that the total data flow sent out by the ground signal station must be equal to the total data flow received by the destination ground signal station. Formula (4.37) constrains that each relay node can only select one optimal neighbor node as the next-hop target. Formula (4.38) constrains that the physical link can only select to activate one of the modes, FSO or RF, at any given time. Formula (4.39) constrains that the normalized residual energy of the UAV node must be higher than the energy safety threshold preset by the system.

[0061] Problem Analysis In FSO / RF-UANET, existing routing algorithms have problems such as being unable to perceive the unique attributes of different link modes, lacking global energy awareness, being unable to cope with sudden clock drift interference, and having defects in link switching schemes.

[0062] To address the aforementioned issues, this invention proposes a joint optimization algorithm for FSO / RF-UANET routing and localization based on link-state-aware spatiotemporal collaboration (E-MA-HGNN & EST-UL-GNN). The E-MA-HGNN sub-algorithm constructs high-dimensional edge attribute features and proposes an edge-enhanced graph attention mechanism, capable of distinguishing link modes and link quality, and assigning different forwarding weights to different links. Furthermore, it transforms passive rule switching in routing decisions into active mode selection based on deep features. The EST-UL-GNN sub-algorithm achieves flexible optimization of UAV position and synchronization strategies through a spatiotemporal dual-stream architecture. The spatial stream captures node movement and coverage requirements, the temporal stream predicts clock deviation and synchronization risks, and the elastic coupling layer balances spatial position adjustment and synchronization stability. The two sub-algorithms work synergistically to seek superior overall performance.

[0063] To address the aforementioned problems, this invention proposes a joint optimization algorithm for FSO / RF-UANET routing and localization based on link-state-aware spatiotemporal coordination. This algorithm includes a link-state-aware graph neural network routing optimization sub-algorithm to solve the relay path optimization problem, a flexible spatiotemporally synchronized graph neural network localization optimization sub-algorithm to solve the UAV localization optimization problem, and a joint optimization strategy. The algorithm framework diagram is shown below. Figure 3 As shown In FSO / RF-UANET networks, free-space optical communication (FSO) and radio frequency (RF) communication links are introduced to combine the ultra-high bandwidth advantage of FSO links with the wide coverage and high reliability advantages of RF links. However, these two types of links have fundamental differences in their physical properties. FSO links are extremely sensitive to alignment errors and atmospheric turbulence, exhibiting ultra-high bandwidth but poor stability; RF links, on the other hand, are limited by spectrum resources and multipath effects, exhibiting low bandwidth but high stability.

[0064] Traditional routing algorithms abstract links as simple weighted edges, making it difficult to capture the complex physical layer heterogeneity. The E-MA-HGNN sub-algorithm aims to model edge features through graph neural networks, enabling the perception and intelligent switching of the physical state of heterogeneous links. It explicitly incorporates edge features into the entire process of message passing and attention calculation, allowing the network to perceive the physical quality of the links. Mathematically, this involves extending graph convolution operations from... Expand to .

[0065] The E-MA-HGNN sub-algorithm inherits the reinforcement learning modeling framework of the HA-GNN sub-algorithm in its routing decision-making process, but updates the original single distance reward. A comprehensive evaluation function is embedded into the underlying feedback mechanism of reinforcement learning; when the agent successfully constructs a complete relay path... Then, the comprehensive evaluation function will assess the overall quality of the entire path and use the evaluation score as the global delay reward. Return it to the policy network.

[0066] The input features and message passing mechanism will be described in detail below.

[0067] Physical modeling of node features In FSO / RF-UANET, the high dynamic characteristics, energy-constrained nature, and heterogeneity of UAV nodes must be considered. Therefore, these physical conditions must be transformed into multi-dimensional characteristic attributes of the graph nodes. Any node in the network... In the time slot The dynamic physical characteristic states are modeled as high-dimensional feature tensors. It includes the following key physical quantities: Three-dimensional spatial position representation : Maneuvering inertial characteristics : The instantaneous velocity of the drone node in each coordinate axis direction, .

[0068] Node normalized residual energy characteristics : This refers to the remaining energy at each node, using the initial energy. Normalize it. .

[0069] Node role identifier characteristics One-Hot encoding is used to distinguish the roles of nodes (ground signal stations and drone nodes).

[0070] Physical modeling of edge features When constructing the input to a graph neural network, we not only define the node feature matrix but also construct the edge feature tensor. Each edge... eigenvectors It includes the following key physical quantities: Geometric distance features : This is the fundamental variable that determines the path loss in free space.

[0071] Jitter variance prediction : .

[0072] Historical channel status Used to record the past The signal-to-noise ratio of each time slot is a key feature for verifying link reliability.

[0073] Link mode identifier Used to distinguish between FSO links and RF links.

[0074] Link Scenario Identifier One-Hot coding is used to distinguish link types, including air-to-air links and air-to-ground links, because different types of links have different divergence angles and receiver aperture parameters.

[0075] Link State Aware Attention Mechanism The HA-GNN proposed in this invention employs a node-centric attention mechanism, and its attention coefficients... Based solely on node features The decision was made, but this attention mechanism is not suitable for FSO / RF-UANET networks. The E-MA-HGNN algorithm introduces an edge fusion layer, enabling the attention mechanism to focus on the heterogeneous properties of the links.

[0076] Since FSO and RF have different physical dimensions, in order to map heterogeneous physical quantities to the same semantic space, node and edge features are first projected using a learnable linear transformation matrix: (4.40) (4.41) in, Indicates different link types.

[0077] Traditional attention scoring is In the E-MA-HGNN sub-algorithm, the source node, target node, and the edge features between them are concatenated into ternary pairs to construct the attention coefficients. : (4.42) in, This indicates a feature concatenation operation. Indicates the first One's attention. When When dealing with a single FSO link and high jitter, the E-MA-HGNN sub-algorithm will learn through training and assign a larger negative weight to this feature, thereby reducing jitter. This will allow routing to avoid selecting high-bandwidth but extremely unstable FSO links, thus intelligently switching to RF links. And when... Corresponding to one RF link and distance When the frequency is high, path loss becomes a critical factor. The E-MA-HGNN sub-algorithm reduces its weight, and routing will avoid selecting long-distance, low-speed RF links. For different networks, network quality is multi-dimensional; sometimes low latency is required, and sometimes high jitter is needed. To capture these characteristics simultaneously, this invention employs... Each attention head has its own independent attention head, and different attention heads focus on different features. The first attention head may focus on distance features, while the second attention head may focus on distance features. An individual's attention may focus on jitter features.

[0078] Furthermore, when aggregating neighbor information, we not only transmit the state of neighbor nodes... Furthermore, the characteristics of the link itself are also superimposed. To more precisely control the impact of edge features on node information, a fusion formula based on a gating mechanism is used. First, a fusion gating coefficient is calculated. It is determined by both node features and edge features: (4.43) in, It is the Sigmoid activation function, and its output range is... .

[0079] Then, the node features and edge features are weighted and mixed using this gating coefficient: (4.44) in, It represents the Hadamah accumulation. From node Passed to node The news. It is considered a learnable gating filter. If the link quality is extremely poor, the E-MA-HGNN sub-algorithm... Continuous learning, approaching 0, at this point the message... Will be mainly composed of edge features Dominate, and block nodes Specific status Conversely, if the link quality is ideal, then node information is mainly transmitted.

[0080] Ultimately, the node Aggregate messages from all neighbors and update its own feature representation. : (4.45) in, It is a non-linear activation function.

[0081] This mechanism ensures the overall network communication quality; if a link has extremely poor quality, it will not only reduce the attention coefficient... To reduce the impact, the message it conveys It also carries the characteristic of poor link quality. (Receiving node) Upon receiving this message, the system will recognize that the neighbor in that direction is unreachable in the next layer of decision-making, making the entire network aware of the deterioration in the communication quality of that area.

[0082] Policy gradient training Define the objective function of the E-MA-HGNN sub-algorithm as the policy. Expected value of the expected reward: (4.46) in, This represents a complete relay path. This is the reward corresponding to that path.

[0083] Since routing decisions are discrete actions, it's impossible to directly calculate the mathematical derivative of the action selected for the discrete next-hop node. Therefore, this invention uses the standard logarithmic derivative to generate the policy gradient. The agent is in state... Below, the decoder outputs the probability of selecting a specific action. Recorded as .

[0084] Used to optimize the objective function parameters The policy gradient can be described as (4.47) in, This represents the total number of hops for this relay path. This refers to the state of the decoder network. Select node for output The probability of.

[0085] Therefore, E-MA-GNN The parameters can be updated as follows: (4.48) in, The learning rate is the value of the E-MA-GNN sub-algorithm.

[0086] After using the E-MA-HGNN sub-algorithm to solve the sub-problem of how to select the optimal relay path, this invention will use the EST-UL-GNN sub-algorithm to solve the sub-problem of UAV positioning and synchronization optimization.

[0087] Elastic spatiotemporal coupling In FSO link communication, time synchronization accuracy directly determines the success rate of signal reception. Traditional fixed-period synchronization mechanisms are based on the assumption that environmental changes are linear and clock drift is predictable. However, when the UAV moves at high speed or the ambient temperature changes drastically, it cannot cope with the nonlinear drift of the crystal oscillator, leading to time slot misalignment and causing FSO link interruption. In the EST-UL-GNN sub-algorithm, it is proposed to treat the UAV's spatial position and clock synchronization state as a strongly coupled pair of variables. By sacrificing some spatial degrees of freedom, the UAV node is moved to a more advantageous position, thereby improving the stability of time synchronization. This also maximizes the spatial coverage and increases spatial degrees of freedom when the clock is stable.

[0088] In dynamic FSO / RF-UANET networks, system robustness depends on spatial, temporal, and environmental disturbances. This invention defines the system state as a composite vector. .

[0089] spatial state The spatial state of the UAV node is a key factor determining the geometric gain and propagation delay of the FSO link. Define the first... The spatial state of each node is: (4.49) in, These represent the position, instantaneous velocity, and acceleration of the UAV node in a three-dimensional Cartesian coordinate system, respectively.

[0090] Time Status The nonlinear drift of the airborne crystal oscillator determines the alignment accuracy of the time slots. The deviation of the local clock relative to the system reference time can be modeled as: (4.50) in, Represents the initial phase deviation. Represents the normalized frequency deviation. This represents the frequency aging drift rate.

[0091] In continuous time, clock skew can be modeled as: (4.51) in, This represents a random process composed of phase noise and white noise. It is a temperature-sensitive higher-order term, which enables the time stream to capture the nonlinear impact on frequency stability caused by the rapid temperature drop due to the rapid ascent of the drone.

[0092] Environmental conditions Environmental conditions include temperature Wind speed It can be defined as: (4.52) The EST-UL-GNN algorithm learns a depth mapping function. The optimal response strategy is inferred in real time. The mapping function can be defined as: (4.53) This function outputs the displacement vector of the drone node at the next moment. Furthermore, it evaluates the probability of current clock state deterioration and outputs a probability value to determine when to initiate the global synchronization process.

[0093] Dual-stream LSTM network architecture The EST-UL-GNN algorithm adopts a learning architecture based on a dual-stream structure, which includes a spatial stream and a temporal stream, and the two interact deeply on features.

[0094] Space Flow Spatial flow is used to model the physical motion and network topology of UAVs. Its input mainly consists of multi-dimensional time-series inputs, including historical trajectory sequences. And global topological embeddings extracted by the preorder graph neural network Unlike traditional unidirectional LSTM, Spatial Flow uses a bidirectional long short-term memory network (Bidirectional LSTM) as its core encoder. It can simultaneously capture the inertial laws of UAV motion and the constraints of target position. It can use past positions and combine them with neighbor position information extracted by graph neural networks to predict the ideal position that maximizes network connectivity.

[0095] The state update formula is: (4.54) in, yes The set of all internal weight matrices, eigenvectors This indicates the spatial requirements for drone nodes to maintain optimal link connectivity in the future.

[0096] Time Flow The time stream is used to handle the hidden variables of physical layer clock evolution, primarily the physical characteristics of the crystal oscillator. Its inputs include past clock skew. With key environmental parameters, the input vector is Because clock drift exhibits strong unidirectional causality, the time flow will be modeled using a unidirectional LSTM. (4.55) in, It is the set of weight parameters for a unidirectional LSTM.

[0097] Temperature is a major factor affecting the frequency stability of crystal oscillators. LSTM can predict future clock deviation trends by learning the correlation between temperature change curves and clock drift curves. When it is predicted that the crystal oscillator is about to enter the high-frequency drift region, this characteristic will fluctuate strongly, indicating an increase in risk.

[0098] Elastic coupling layer The flexible coupling layer is the coordinating unit, responsible for balancing spatial and temporal requirements. It weighted and fused the characteristics of spatial and temporal flows: (4.56) in, It is a linear projection layer used to align spatial feature dimensions; It is the elasticity coefficient, used to adjust the system's sensitivity to spatial movement; It provides a learnable reference offset. When the ambient temperature is constant, the clock drift is minimal. The activation value is low and stable. At this time... The gradient is mainly affected by The impact will be prioritized for coverage area, with drone nodes deployed in a distributed manner to maximize network coverage. However, when LSTM predicts an impending severe clock drift, The enhanced impact of this makes the coupling features more focused on synchronization stability. Subsequently, the model calculates the gradient based on the unsupervised loss function and, through backpropagation, forces the UAV to sacrifice its current spatial advantage and actively move to an area with better link quality or less interference in order to obtain a more accurate synchronization signal and prevent link interruption.

[0099] Unsupervised loss function This invention designs a multi-objective unsupervised loss function. This loss function transforms the physical laws of link communication into optimization constraints, and the EST-UL-GNN algorithm explores the spatiotemporal equilibrium point through self-evolution. Loss Function Defined as a weighted sum across four dimensions: (4.57) in, The parameters are used to adjust the weights of each item.

[0100] Coverage loss This is used to minimize the sum of squared distances from all ground signal stations to the drone node, ensuring basic geometric coverage: (4.58) in, It is a ground signal station coordinates It is the position adjustment amount output by the model. This enables drone swarms to cover areas with dense ground signal stations.

[0101] Link quality loss Link quality loss is used to maximize the average signal-to-noise ratio of the entire network. The edge features calculated by the E-MA-HGNN sub-algorithm are used to maximize the average link score of the entire network.

[0102] (4.59) Synchronous risk loss To prevent time slot misalignment caused by nonlinear drift of the crystal oscillator, this invention introduces an exponential penalty term: (4.60) in, This represents the clock skewness of future moments predicted by the time-flow LSTM. This indicates the maximum allowable gate deviation threshold of the system. It is a decay coefficient used to adjust the sensitivity to the risk of synchronization loss.

[0103] Communication power consumption penalty To meet the physical constraints of the drone's limited energy, limit its maximum travel distance, and prevent battery depletion due to frequent large-scale maneuvers, a displacement penalty term is introduced: (4.61) In order to continuously adjust its network parameters To output better coordinates, it is necessary to calculate the gradient of the objective function with respect to the weights of the deeper layers of the neural network. According to formula (4.53), the total unsupervised loss function for multi-objective joint optimization is: (4.62) Finally, EST-UL-GNN utilizes an unsupervised learning rate. Perform gradient descent update operation on the parameters, as shown in the formula. (4.63) Joint training mechanism of two models Phase 1: E-MA-HGNN Routing Decision Training. Training samples are constructed based on the generated RF / FSO heterogeneous topology. Each sample contains features such as node location, remaining node energy, link distance, link mode, link quality, jitter variance, and historical channel state. E-MA-HGNN extracts the joint representation of nodes and links through edge-enhanced graph attention, and the decoder outputs the next-hop selection probability. After the agent constructs a complete path, the path reward is calculated. The network parameters are updated using policy gradients, allowing the model to gradually learn the active mode selection strategy under different RF / FSO link states. After training convergence, the E-MA-HGNN weight parameters are solidified, transforming it into an evaluator with evaluation capabilities.

[0104] Phase Two: EST-UL-GNN Spatiotemporal Localization Optimization Training. In this phase, the E-MA-HGNN model with frozen weights is loaded. Based on the dual-stream LSTM network of the EST-UL-GNN, the model outputs a new spatial displacement vector. Subsequently, link distance, link mode, link quality, and synchronization risk are recalculated based on the updated location, and the RF / FSO heterogeneous graph is reconstructed. The updated heterogeneous graph is input into the frozen E-MA-HGNN, and the unsupervised loss is calculated. By updating parameters through backpropagation, the EST-UL-GNN can learn location adjustment directions that are beneficial for improving routing quality and reducing synchronization risk.

[0105] Collaborative simulation State perception and heterogeneous graph reconstruction During the current operational phase, the system collects network status data, including the UAV's 3D position, speed, remaining energy, link distance, link mode, link quality, jitter variance, clock skew, etc., and reconstructs this data into the current heterogeneous graph.

[0106] EST-UL-GNN Elastic Spatiotemporal Collaboration The drone's position adjustment vector is output based on the current spatiotemporal state. If the clock is predicted to be in a high-risk period of high-frequency drift, the spatial stream will actively output maneuver commands that sacrifice some coverage area to guide the drone toward a safer area with less interference or a better link, in exchange for a high-precision synchronization signal; if the clock state is stable, the focus will be on coverage and link quality to form a high-quality transmission path.

[0107] E-MA-HGNN Active Mode Selection and Routing Decision As the physical location of the drone is slightly adjusted, the network topology changes. E-MA-HGNN performs route reasoning based on the updated graph topology, calculates edge-enhanced attention weights based on node states and edge attributes, and outputs the optimal relay path.

[0108] Through the above loop, the algorithm exhibits superior overall performance across multiple performance metrics.

[0109] The pseudocode for the E-MA-HGNN & EST-UL-GNN joint optimization algorithm is shown below.

[0110] The simulation experiments and performance comparisons of this invention include the following: To verify the effectiveness of the E-MA-HGNN & EST-UL-GNN joint optimization algorithm in the FSO / RF-UANET scenario, this invention further introduces mechanisms such as the FSO / RF dual-mode link model, link mode selection, node remaining energy update, and network fault interference. In comparative experiments, the proposed E-MA-HGNN & EST-UL-GNN joint optimization algorithm is compared with the proposed HA-GNN & UL-GNN joint optimization algorithm, and its performance is compared with the AODV algorithm, Q-Routing algorithm, and a two-stage algorithm consisting of Bellman-Ford algorithm and MLP. To consider the relative performance differences of different routing decision mechanisms under unified topology, link model, and service load conditions, this invention implements AODV, Q-Routing, and BF-MLP as algorithm-level baselines. The AODV baseline retains the core ideas of on-demand routing and shortest hop count priority; the Q-Routing baseline retains the hop-by-hop routing learning mechanism based on Q-value estimation; and BF-MLP represents the baseline based on Bellman-Ford path search and multilayer perceptron cost prediction. This comparison includes a single-stage routing control group, a two-stage routing control group, and the HA-GNN & UL-GNN joint optimization algorithm proposed in this invention, comprehensively evaluating the superiority of the E-MA-HGNN & EST-UL-GNN joint optimization algorithm. Furthermore, to further verify the effectiveness of each sub-algorithm of E-MA-HGNN & EST-UL-GNN, this invention conducted module ablation experiments with varying node sizes under the same network topology, communication radius, and link model conditions. The removed E-MA-HGNN and EST-UL-GNN were replaced by Greedy greedy routing and KMeans, respectively.

[0111] The simulation environment setup mainly includes network configuration, heterogeneous communication model configuration, and parameter configuration for the E-MA-HGNN & EST-UL-GNN joint optimization algorithm. Table 4.1 lists the basic network topology configuration.

[0112] Table 4.1 Network Configuration Parameter Table

[0113] The communication model configuration covers the simulation parameters of FSO / RF heterogeneous channels, as shown in Table 4.2.

[0114] Table 4.2 Simulation Parameters for FSO / RF Heterogeneous Channel

[0115] The core parameter settings for the E-MA-HGNN & EST-UL-GNN joint optimization algorithm are shown in Table 4.3.

[0116] Table 4.3 Simulation Parameter Table

[0117] Results Analysis (1) End-to-end delay Figure 4The figure compares the end-to-end latency of the E-MA-HGNN & EST-UL-GNN algorithm with four other algorithms at different node scales. The figure shows that the E-MA-HGNN & EST-UL-GNN algorithm significantly outperforms the other three algorithms in terms of end-to-end latency across different numbers of nodes. As the number of nodes increases, the E-MA-HGNN & EST-UL-GNN algorithm maintains low latency and continues to decrease. This is because with a larger node scale, E-MA-HGNN can accurately identify and select high-quality FSO links, while EST-UL-GNN uses dual-stream LSTM for location optimization. The combined effect of these two components reduces the selection of low-quality links, thus maintaining low latency. The HA-GNN & UL-GNN algorithm has a latency of approximately 0.075 seconds to 0.080 seconds, which is inferior to the E-MA-HGNN & EST-UL-GNN algorithm. This algorithm does not explicitly model the differences between FSO / RF dual-mode links, and cannot distinguish between FSO and RF links. Its adaptability to different link modes is weaker than E-MA-HGNN and EST-UL-GNN. AODV exhibits significant linear degradation, with its latency increasing linearly with the number of nodes. When the network size is N20-M2, its latency reaches 0.083 seconds, while when the network size reaches N60-M10, the latency reaches 0.101 seconds. As the number of nodes increases, on-demand path search needs to check more nodes and edges. Simultaneously, paths with fewer hops do not necessarily have higher link capacity; AODV may select paths with higher transmission costs, leading to an increase in end-to-end latency as the network size expands. BF-MLP and Q-Routing are relatively stable, with better latency performance than AODV. As the number of nodes increases, the latency of BF-MLP slowly increases from 0.073 seconds to 0.079 seconds. This algorithm utilizes MLP to avoid high-cost links to some extent, but MLP lacks the ability to deeply model the states of FSO / RF heterogeneous links. Therefore, as the network scales up and the link combinations become more complex, its path cost estimation is prone to errors, and the latency increases accordingly. Q-Routing mainly updates the Q value based on local historical feedback, and its ability to comprehensively perceive global topology, link mode differences, and multi-hop path bottlenecks is limited.

[0118] (2) Package delivery rate Figure 5The graph compares the packet delivery rates of the E-MA-HGNN & EST-UL-GNN algorithms with four other algorithms at different node scales. As shown in the graph, the E-MA-HGNN & EST-UL-GNN algorithms have a higher packet delivery rate than the other four, and this rate steadily increases with the node scale, ultimately reaching 97.22% at a node scale of N60-M10. This is because E-MA-HGNN can perform reliability assessments on heterogeneous FSO / RF links. The algorithm reduces the weight of low-quality or unstable links in path selection, thereby reducing transmission failures caused by insufficient link quality. EST-UL-GNN further improves the transmission conditions of some links by adjusting the UAV's position and synchronization status through spatiotemporal coordination, thus maximizing the packet delivery rate. The packet delivery rate of the HA-GNN & UL-GNN algorithms remains around 0.93, only lower than that of the E-MA-HGNN & EST-UL-GNN algorithms. This algorithm can utilize the current topology, node status, and link status for path selection and improve FSO link quality through position optimization. BF-MLP has the lowest packet delivery rate. In environments with varying conditions, its judgment of path reliability may be inaccurate. If a link that does not meet the transmission conditions is selected, the data packet will be judged as a transmission failure, resulting in a large number of packet losses. AODV performs well with a small number of nodes, reaching 93.22%, but its packet delivery rate decreases continuously as the number of network nodes increases. This is because AODV can find feasible paths quickly with a small number of nodes, resulting in a higher packet delivery rate. As the number of nodes increases, topology changes and the number of candidate paths increase simultaneously, making on-demand path search more susceptible to changes in link status. Some paths may fail after the search is completed due to deterioration in link quality, leading to a decrease in packet delivery rate. Q-Routing maintains a packet delivery rate of around 91%. It has some adaptive capability, but it mainly relies on the update of the Q-Table. When the update of the Q-Table lags behind topology changes, packet loss will occur.

[0119] (3) Network throughput Figure 6The simulation results demonstrate a comparison of network throughput between the E-MA-HGNN & EST-UL-GNN algorithms and four other algorithms at different node scales. Simulation results show that the E-MA-HGNN & EST-UL-GNN algorithms maintain the highest throughput at larger node scales, reaching 16.926 Mbps. This is because E-MA-HGNN introduces a link-state-aware attention mechanism, explicitly distinguishing between RF and FSO links. Since FSO link bandwidth is much higher than RF link bandwidth, the algorithm assigns higher weight to FSO links through the attention mechanism. EST-UL-GNN uses dual-stream LSTM to adjust the UAV's position, improving the quality of FSO links and ensuring that FSO links are maintained for transmission as much as possible while avoiding RF links. Q-Routing performs second best and is relatively stable, maintaining between 14-15 Mbps. It dynamically learns network congestion conditions by maintaining a Q-Table, which can avoid historically high-cost links to some extent and improve forwarding performance using local experience, achieving local load balancing. However, it only updates the Q value based on feedback and cannot distinguish link modes, resulting in a slightly lower throughput than the E-MA-HGNN & EST-UL-GNN algorithms in high-dynamic scenarios. The HA-GNN & UL-GNN algorithms are slightly worse than Q-Routing, maintaining a throughput between 14-14.5 Mbps. HA-GNN & UL-GNN can improve FSO link quality by utilizing heterogeneous attention mechanisms and location optimization, but they lack the ability to distinguish between different link modes. When low-capacity RF links or unstable FSO links appear in the path, the effective throughput will decrease. AODV's throughput decreases sharply with the increase in node size because on-demand path search may select paths with fewer hops but lower link capacity or insufficient stability. If the path contains many low-capacity RF links, or some FSO links do not meet the transmission conditions, the amount of data successfully received by the destination node will decrease, thereby reducing the effective received data volume and ultimately leading to a decrease in network throughput.

[0120] (4) Energy consumption of end-to-end communication Figure 7This paper compares the end-to-end communication energy consumption of the E-MA-HGNN & EST-UL-GNN algorithms with four other algorithms under different node numbers. As the number of nodes increases, the end-to-end communication energy consumption of the proposed E-MA-HGNN & EST-UL-GNN algorithms decreases significantly, from 1435.03 mJ to 1079.08 mJ. This is because as the number of selectable relay nodes in the network increases, the algorithm can select lower-cost paths from more relay nodes and links. The loss function of EST-UL-GNN includes an explicit energy penalty term, improving the connectivity of some links through location adjustments. The end-to-end communication energy consumption of the HA-GNN & UL-GNN algorithms is generally around 1400 mJ, without showing a significant decreasing trend across different node sizes. The algorithms do not explicitly distinguish between FSO and RF links, nor do they model the differences in communication energy consumption under different link modes. Higher communication energy consumption occurs when RF backoff links or high-cost FSO links exist in the network. The baseline communication energy consumption of the three algorithms, AODV, Q-Routing, and BF-MLP, is stable and does not change significantly with the increase in the number of nodes. However, when the number of nodes increases, they all lack an assessment of the difference in communication energy consumption between FSO and RF links, and cannot make full use of the low-energy communication paths provided by large-scale topologies, thus failing to effectively reduce end-to-end communication energy consumption.

[0121] (5) Mode switching quantity Figure 8The graph illustrates the changes in mode switching frequency between the E-MA-HGNN & EST-UL-GNN algorithms and other algorithms. Mode switching frequency primarily describes the degree of change in the communication mode of a link within a path between adjacent time slots, i.e., whether the link frequently switches between RF and FSO links. A smaller value indicates a more stable path mode and lower routing switching overhead. As shown in the graph, the mode switching frequency of the E-MA-HGNN & EST-UL-GNN algorithms remains consistently low, with only slight fluctuations in a few time slots. The E-MA-HGNN sub-algorithm uses a multi-head attention mechanism to distinguish between FSO and RF links, enabling the algorithm to identify them. The EST-UL-GNN sub-algorithm uses dual-stream LSTM to predict UAV position changes and clock drift trends, avoiding the selection of high-jitter links or links with synchronization risks. The combined effect of these two algorithms makes the algorithm more inclined to select paths with stable modes and good link quality continuity, thus maintaining a low level of mode switching frequency. The BF-MLP algorithm exhibits the highest mode switching frequency and significant fluctuations. While BF-MLP can estimate path costs to some extent using MLP, it primarily relies on static feature mapping and lacks explicit modeling of heterogeneous FSO / RF link modes. When link states are affected by UAV movement, FSO alignment disturbances, or channel fluctuations, BF-MLP tends to select links with different modes in adjacent time intervals, leading to frequent switching between FSO and RF. AODV has a lower overall mode switching rate, but this does not necessarily indicate strong heterogeneous link awareness. AODV path selection is mainly based on topological connectivity and hop count, and it does not actively perceive differences in FSO / RF link modes. The lower AODV mode switching rate is mainly due to its relatively conservative path selection and lack of a mechanism for actively adjusting link modes. However, when the network topology changes, AODV needs to re-initiate the route discovery process, which may lead to mode changes. Q-Routing has a lower mode switching rate than BF-MLP, but still exhibits significant switching peaks in certain time periods. Q-Routing mainly relies on local historical feedback to update the Q value, which can avoid links with poor historical performance to a certain extent. However, its state representation is usually difficult to fully describe the FSO / RF link modes, channel jitter and synchronization risks, and it has lag, which leads to an increase in mode switching in some time slots.

[0122] (6) Throughput recovery trend under sudden node failure scenario Figure 9This study compares the throughput recovery trends of the E-MA-HGNN & EST-UL-GNN algorithms with three other baseline algorithms after a sudden failure of some nodes in an N60-M10 network. To simulate a sudden situation, the simulation system destroyed some active nodes in the network at 200 seconds. The E-MA-HGNN & EST-UL-GNN algorithm dropped from approximately 17.5 Mbps to approximately 14 Mbps, then gradually recovered and stabilized at around 15 Mbps. This is because the E-MA-HGNN & EST-UL-GNN algorithms employ a message-passing neural network mechanism, where nodes continuously exchange high-dimensional feature vectors. When a critical relay node fails, neighboring nodes can quickly detect this and rapidly transmit this state change to the entire UAV network through updated graph features, rebuilding the network topology. Simultaneously, the spatiotemporal coordination mechanism of the EST-UL-GNN sub-algorithm also helps improve the connection conditions of some links, thereby increasing the effective throughput after the failure. After the AODV encountered a failure at 200 seconds, its throughput experienced a precipitous drop. AODV has weak perception of FSO / RF link modes, link capacity, and link stability. When the original path is affected by node failure, the baseline needs to find a feasible path again in the remaining topology, and the re-formed path may still contain links with low capacity or insufficient stability, resulting in lower throughput after recovery. BF-MLP's throughput drops to about 6 Mbps, and then gradually recovers to about 8 Mbps. When node failure causes changes in the set of available links, its link estimation will lag and be biased, leading to the selection of low-quality paths after the failure. Q-Routing's throughput drops to about 8.5 Mbps, and gradually recovers to about 10-11 Mbps. It relies on local Q-value estimation for hop-by-hop path selection and has a certain degree of adaptability. However, after a sudden node failure, it needs a certain amount of time to readjust the next hop selection, so the recovery speed is lower than that of the algorithm in this invention.

[0123] (7) Comprehensive evaluation of link status Figure 10The results of the comprehensive evaluation of link status under different node sizes are presented. The comprehensive evaluation index considers factors such as link distance, path reliability, bottleneck capacity, energy penalty, and synchronization risk to reflect the link quality of paths in FSO / RF-UANET. The higher the index, the better the overall performance of the algorithm in terms of heterogeneous link selection, path stability, and spatiotemporal coordination. As shown in the figure, the E-MA-HGNN & EST-UL-GNN algorithms achieve the highest comprehensive evaluation value at any node size. When the network size increases from N20-M2 to N60-M10, the comprehensive evaluation value of the algorithm shows an upward trend, increasing from approximately 0.49 to approximately 0.64. As the number of UAV nodes in the network increases, the algorithm can obtain more link sets and perceive the status of FSO / RF heterogeneous links through the E-MA-HGNN sub-algorithm, thereby prioritizing links with higher capacity and better stability. At the same time, EST-UL-GNN can optimize and adjust the UAV position according to the spatiotemporal status, so that the topology brought by the new nodes can be transformed into effective link quality gains. Removing the EST-UL-GNN sub-algorithm, the overall evaluation value of using EST-UL-GNN without EST-UL-GNN is lower than that of the full algorithm at all node scales. This indicates that relying solely on E-MA-HGNN for link-state aware routing, while able to identify relatively good FSO / RF links, lacks the flexibility for spatiotemporal positioning adjustments, preventing UAV nodes from further improving link quality. In the N50-M8 and N60-M10 scenarios, the gap between using EST-UL-GNN without EST-UL-GNN and the full algorithm narrows. This suggests that when the node scale is large, the topology itself already provides many optional relay paths, and the gains from positioning optimization are relatively weakened. After removing the E-MA-HGNN sub-algorithm, the algorithm's overall evaluation value remains at a low level, and the improvement is not significant with increasing node scale. This phenomenon indicates that with the E-MA-HGNN sub-algorithm, the algorithm cannot fully distinguish the differences between RF and FSO links, and it is difficult to utilize the high capacity advantage of FSO links. Even with increased network scale, newly added relay nodes cannot be effectively utilized, resulting in an overall evaluation value that is consistently lower than that of the full algorithm and without EST-UL-GNN. Overall, the E-MA-HGNN sub-algorithm enhances the routing decision-making's ability to perceive heterogeneous links by considering factors such as link mode, channel quality, and energy state. The EST-UL-GNN sub-algorithm further improves link stability and spatiotemporal coordination through spatial location adjustment and synchronization risk prediction. Both play important roles, enabling the algorithm to exhibit good scalability and comprehensive optimization capabilities in FSO / RF heterogeneous ad hoc network scenarios.

Claims

1. A joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination, characterized in that, The method includes the following steps: Step 1: Physical modeling of node features; In FSO / RF-UANET, it is necessary to consider the high dynamic characteristics, energy-constrained characteristics, and heterogeneity of UAV nodes, and to transform these physical conditions into multi-dimensional feature attributes of graph nodes. Any node in the network... In the time slot The dynamic physical characteristic states are modeled as high-dimensional feature tensors. ; Step 1-1: Three-dimensional spatial position representation : ; Step 1-2: Maneuvering Inertial Characteristics : The instantaneous velocity of the drone node in each coordinate axis direction, ; Steps 1-3: Node Normalized Residual Energy Characteristics : This refers to the remaining energy at each node, using the initial energy. Normalize it. ; Steps 1-4: Node Role Identifiers One-Hot encoding is used to distinguish the roles of nodes (ground signal stations and drone nodes). Step 2: Physical modeling of edge features; When constructing the input to a graph neural network, not only are node feature matrices defined, but edge feature tensors are also constructed, with each edge... eigenvectors It includes the following key physical quantities; Step 2-1: Geometric distance features : This is the fundamental variable that determines the path loss in free space; Step 2-2: Jitter Variance Prediction : ; Steps 2-3: Historical Channel States Used to record the past The signal-to-noise ratio of each time slot serves as a key characteristic for verifying link reliability. Steps 2-4: Link Mode Identification Used to distinguish between FSO links and RF links; Steps 2-5: Link Scenario Identification One-Hot coding is used to distinguish link types, including air-to-air links and air-to-ground links, because different types of links have different divergence angles and receiver aperture parameters. Step 3: Link state-aware attention mechanism; HA-GNN employs a node-centric attention mechanism, and its attention coefficient... Based solely on node features The decision was made, but this attention mechanism is not suitable for FSO / RF-UANET networks. The E-MA-HGNN algorithm introduces an edge fusion layer, which enables the attention mechanism to focus on the heterogeneous properties of the links. First, the node and edge features are projected using a learnable linear transformation matrix: (4.40) (4.41) in, Indicates different link types, In the E-MA-HGNN sub-algorithm, the source node, target node, and the edge features between them are concatenated into ternary pairs to construct the attention coefficients. : (4.42) in, This indicates a feature concatenation operation. Indicates the first One's attention, when When dealing with a single FSO link and high jitter, the E-MA-HGNN sub-algorithm will learn through training and assign a larger negative weight to this feature, thereby reducing jitter. This will allow the router to avoid selecting high-bandwidth but extremely unstable FSO links, thus intelligently switching to RF links, and when Corresponding to one RF link and distance When the path loss is large, it becomes a critical factor. The E-MA-HGNN sub-algorithm reduces its weight, and routing will avoid selecting long-distance, low-speed RF links. For different networks, network quality is measured in multiple dimensions; sometimes low latency is required, and sometimes high jitter is needed. To capture these characteristics simultaneously, a method is adopted... Each attention head has its own independent attention head, and different attention heads focus on different features. The first attention head may focus on distance features, while the second attention head may focus on distance features. An individual's attention may focus on jitter features; Step 4: Policy gradient training; Define the objective function of the E-MA-HGNN sub-algorithm as the policy. Expected value of the expected reward: (4.46) in, This represents a complete relay path. This is the reward corresponding to that path; Since routing decisions are discrete actions, it's impossible to directly calculate the mathematical derivative for selecting actions at discrete next-hop nodes. Therefore, the standard logarithmic derivative is used to generate the policy gradient, and the agent is in the state... Below, the decoder outputs the probability of selecting a specific action. Recorded as ; Used to optimize the objective function parameters The policy gradient is described as follows: (4.47) in, This represents the total number of hops for this relay path. This refers to the state of the decoder network. Select node for output The probability of; Therefore, E-MA-GNN The parameters can be updated as follows: (4.48) in, The learning rate is the value of the E-MA-GNN sub-algorithm. After using the E-MA-HGNN sub-algorithm to solve the sub-problem of how to select the optimal relay path, the EST-UL-GNN sub-algorithm is used to solve the UAV positioning and synchronization optimization sub-problem. Step 5: Elastic spatiotemporal coupling; To prevent FSO link interruption, the EST-UL-GNN sub-algorithm proposes to treat the UAV's spatial position and clock synchronization state as a pair of strongly coupled variables. By sacrificing some spatial degrees of freedom, the UAV node is moved to a more advantageous position to achieve stability of time synchronization. It can also maximize spatial coverage and increase spatial degrees of freedom when the clock is stable. The method defines the system state as a composite vector. ; spatial state The spatial state of the UAV node is a key factor determining the geometric gain and propagation delay of the FSO link. The first... The spatial state of each node is: (4.49) in, These represent the position, instantaneous velocity, and acceleration of the UAV node in a three-dimensional Cartesian coordinate system, respectively. Step 6: Dual-stream LSTM network architecture; The EST-UL-GNN algorithm employs a learning architecture based on a dual-stream structure, which includes a spatial stream and a temporal stream, with the two interacting deeply on features. Spatial flow is used to model the physical motion and network topology of UAVs. Its input mainly consists of multi-dimensional time-series inputs, including historical trajectory sequences. And global topological embeddings extracted by the preorder graph neural network Unlike traditional unidirectional LSTM, Spatial Flow uses a bidirectional long short-term memory network (Bidirectional LSTM) as its core encoder, which can simultaneously capture the inertial laws of UAV motion and the constraints of target position. It can use past positions and combine them with neighbor position information extracted by graph neural networks to predict the ideal position that maximizes network connectivity. The state update formula is: (4.54) in, yes The set of all internal weight matrices, eigenvectors This indicates the spatial requirements for drone nodes to maintain optimal link connectivity in the future. The time stream is used to handle the hidden variables of physical layer clock evolution, mainly the physical characteristics of the crystal oscillator. Its inputs include past clock offsets. With key environmental parameters, the input vector is Because clock drift has strong unidirectional causality, the time flow will be modeled using a unidirectional LSTM: (4.55) in, It is the set of weight parameters for a unidirectional LSTM. Step 7: Design a multi-objective unsupervised loss function; The loss function transforms the physical laws of link communication into optimization constraints. The EST-UL-GNN algorithm, through self-deduction, explores the spatiotemporal equilibrium point, and the loss function... Defined as a weighted sum across four dimensions: (4.57) in, To adjust the parameters of each weight, This is used to minimize the sum of squared distances from all ground signal stations to the drone node, ensuring basic geometric coverage: (4.58) in, It is a ground signal station coordinates It is the position adjustment amount output by the model. This allows drone swarms to cover areas with dense ground signal stations. Link quality loss Link quality loss is used to maximize the average signal-to-noise ratio of the entire network. It utilizes the edge features calculated by the E-MA-HGNN sub-algorithm to maximize the average link score of the entire network. (4.59) To prevent time slot misalignment caused by nonlinear drift of the crystal oscillator, an exponential penalty term is introduced: (4.60) in, This represents the clock skewness of future moments predicted by the time-flow LSTM. This indicates the maximum allowable gate deviation threshold of the system. It is a decay coefficient used to adjust the sensitivity to the risk of synchronization loss.

2. The joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination according to claim 1, characterized in that, Step 3 includes: When aggregating neighbor information, not only the state of neighbor nodes is transmitted. Furthermore, the characteristics of the link itself are also superimposed. To more precisely control the impact of edge features on node information, a fusion formula based on a gating mechanism is used. First, a fusion gating coefficient is calculated. It is determined by both node features and edge features: (4.43) in, It is the Sigmoid activation function, and its output range is... ; Then, the node features and edge features are weighted and mixed using this gating coefficient: (4.44) in, It represents the Hadamah accumulation. From node Passed to node The news, Treated as a learnable gated filter, if the link quality is extremely poor, the E-MA-HGNN sub-algorithm makes... Continuous learning, approaching 0, at this point the message... Will be mainly composed of edge features Dominate, and block nodes Specific status Conversely, if the link quality is ideal, then node information is mainly transmitted. Ultimately, the node Aggregate messages from all neighbors and update its own feature representation. : (4.45) in, It is a non-linear activation function. This mechanism ensures the overall network communication quality; if a link has extremely poor quality, it will not only reduce the attention coefficient... To reduce the impact, the message it conveys It also carries the characteristic of poor link quality, and the receiving node Upon receiving this message, the system will recognize that the neighbor in that direction is unreachable in the next layer of decision-making, making the entire network aware of the deterioration in the communication quality of that area.

3. The joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination according to claim 1, characterized in that, Step 5 includes: Time Status The nonlinear drift of the airborne crystal oscillator determines the alignment accuracy of the time slots. The deviation of the local clock relative to the system reference time can be modeled as follows: (4.50) in, Represents the initial phase deviation. Represents the normalized frequency deviation. Represents frequency aging drift rate, In continuous time, clock skew can be modeled as: (4.51) in, This represents a random process composed of phase noise and white noise. It is a temperature-sensitive higher-order term, enabling the time stream to capture the nonlinear impact on frequency stability caused by the rapid temperature drop due to the rapid ascent of the drone. Environmental conditions Environmental conditions include temperature Wind speed Defined as: (4.52) The EST-UL-GNN algorithm learns a depth mapping function. The optimal response strategy can be inferred in real time, and the mapping function can be defined as: (4.53) This function outputs the displacement vector of the drone node at the next moment. In addition, it also outputs a probability value by evaluating the probability of the current clock state deteriorating, which determines when to start the global synchronization process.

4. The joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination according to claim 1, characterized in that, Step 6 includes: Temperature is the main factor affecting the frequency stability of crystal oscillators. By learning the correlation between temperature change curves and clock drift curves, LSTM can predict future clock deviation trends. When it is predicted that the crystal oscillator is about to enter the high-frequency drift region, this feature will fluctuate strongly, indicating an increase in risk. The flexible coupling layer is the coordinating unit, responsible for balancing spatial and temporal requirements. It weighted and fused the characteristics of spatial and temporal flows. (4.56) in, It is a linear projection layer used to align spatial feature dimensions; It is the elasticity coefficient, used to adjust the system's sensitivity to spatial movement; It provides a learnable reference offset that minimizes clock drift when the ambient temperature is constant. The activation value is low and stable at this time. The gradient is mainly affected by The impact will be prioritized for coverage area, with drone nodes deployed in a distributed manner to maximize network area, and when LSTM predicts an impending severe clock drift, The enhanced impact makes the coupling features more focused on synchronization stability. Subsequently, the model calculates the gradient based on the unsupervised loss function and forces the UAV to sacrifice its current spatial position advantage and actively move to an area with better link quality or less interference through backpropagation in order to obtain a higher precision synchronization signal and prevent link interruption.

5. The joint optimization method for FSO / RF-UANET routing and location based on link state-aware spatiotemporal coordination according to claim 1, characterized in that, Step 7 includes: To meet the physical constraints of the drone's limited energy, limit its maximum travel distance, and prevent battery depletion due to frequent large-scale maneuvers, a displacement penalty term is introduced: (4.61) In order to continuously adjust its network parameters To output better coordinates, it is necessary to calculate the gradient of the objective function with respect to the weights of the deeper layers of the neural network. According to formula (4.53), the total unsupervised loss function for multi-objective joint optimization is: (4.62) Finally, EST-UL-GNN utilizes an unsupervised learning rate. Perform gradient descent update operation on the parameters, as shown in the formula. (4.63) Joint training mechanism for two models; Phase 1: E-MA-HGNN routing decision training. Training samples are constructed based on the generated RF / FSO heterogeneous topology. Each sample contains features such as node location, node remaining energy, link distance, link mode, link quality, jitter variance, and historical channel state. E-MA-HGNN extracts the joint representation of nodes and links through edge-enhanced graph attention, and the decoder outputs the next-hop selection probability. After the agent constructs a complete path, the path reward is calculated, and the network parameters are updated using policy gradients, so that the model gradually learns the active mode selection strategy under different RF / FSO link states. After training converges, the E-MA-HGNN weight parameters are solidified, transforming it into an evaluator with evaluation capabilities. Phase Two: EST-UL-GNN Spatiotemporal Localization Optimization Training. In this phase, the E-MA-HGNN model with frozen weights is loaded. Based on the dual-stream LSTM network of EST-UL-GNN, the model outputs a new spatial displacement vector. Subsequently, the link distance, link mode, link quality, and synchronization risk are recalculated based on the updated position, and the RF / FSO heterogeneous graph is reconstructed. The updated heterogeneous graph is input into the fixed E-MA-HGNN, and the unsupervised loss is calculated. The parameters are updated through backpropagation. EST-UL-GNN can learn the position adjustment direction that is beneficial to improving routing quality and reducing synchronization risk. Collaborative simulation; State awareness and heterogeneous graph reconstruction; During the current operational phase, the system collects network status data, including the UAV's 3D position, speed, remaining energy, link distance, link mode, link quality, jitter variance, clock skew, etc., and reconstructs this data into the current heterogeneous graph. EST-UL-GNN elastic spatiotemporal collaboration; Based on the current spatiotemporal state, the drone position adjustment vector is output. If it is predicted that the clock is in a high-risk period of high-frequency drift, the spatial stream will actively output a maneuver command that sacrifices part of the coverage area to guide the drone to a safe area with less interference or better link in exchange for a high-precision synchronization signal. If the clock state is stable, the focus will be on coverage and link quality to form a high-quality transmission path. E-MA-HGNN Active Mode Selection and Routing Decision; As the physical location of the drone is fine-tuned, the network topology changes. E-MA-HGNN performs route reasoning based on the updated graph topology, calculates edge-enhanced attention weights based on node states and edge attributes, and outputs the optimal relay path.