A method for antenna-assisted low-latency air-space transmission
By combining an outer second-order cone programming model with an inner K-stage processing network in a block-based iterative solution architecture, the panel and antenna configuration in UAV-assisted aerospace transmission is optimized, solving the problems of decreased information freshness and excessive computation time, and achieving efficient information transmission and rapid decision-making.
Patent Information
- Application Number
- CN202611578351.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-10
AI Technical Summary
In drone-assisted air-to-ground collaborative transmission scenarios, existing technologies struggle to effectively optimize the configuration of active reconfigurable transmission reflector panels and reconfigurable receiving antenna arrays without significantly increasing power consumption, leading to a decrease in information freshness. Furthermore, existing methods have excessively long computation times, failing to meet the demands of online real-time decision-making.
A block-based iterative solution architecture combining an outer second-order cone programming layer and an inner K-stage processing network is adopted to jointly optimize the channel path reconstruction of an active reconfigurable transmission reflector panel, the port reconstruction of a reconfigurable receiving antenna array, and the trajectory planning of a UAV. By constructing a mixed-integer non-convex optimization problem and training the K-stage processing network with a supervision signal, online updates of parameters and port selection are achieved.
While ensuring the quality of multi-user communication, it reduces the average information freshness of the system and shortens the solution time, avoiding the gain loss caused by individual optimization, and improving the stability of information freshness and computational efficiency.
Smart Images

Figure CN123074533A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication technology, and specifically to an antenna-assisted low-latency air-to-space transmission method. Background Technology
[0002] With the development of mobile communication networks, status update services such as drone inspection, vehicle-to-everything (V2X) communication, and industrial IoT data aggregation are placing increasingly stringent demands on data freshness. Information freshness has become one of the key performance indicators for measuring the service quality of such services. In drone-assisted air-to-ground collaborative transmission scenarios, drone nodes are highly mobile, transmission links involve multiple cascaded connections, and channel states change rapidly over time. These factors lead to drastic fluctuations in the effective channel gain from the ground user to the receiver. Simply increasing transmission power or bandwidth is insufficient to reliably guarantee information freshness under limited energy consumption conditions.
[0003] To improve effective channel quality without significantly increasing power consumption, existing technologies incorporate an active reconfigurable transmission reflector panel beneath the UAV. Each element of this panel integrates a power amplifier and supports dual-mode transmission and reflection, enabling the reconfiguration of the channel transmission path from the transmitting end. This allows ground user data to be uploaded to the base station via a reflection link or to the UAV's onboard receiver via a transmission link. Simultaneously, a reconfigurable receiving antenna array is configured on the base station side. This array can dynamically select a subset of active ports within a compact aperture, thereby obtaining spatial diversity gain from the receiving end. In multi-user uplink transmission scenarios, to further enhance system capacity, non-orthogonal multiple access technology combined with serial interference cancellation mechanisms is typically employed.
[0004] However, in practical applications, the configuration optimization of active reconfigurable transmit reflector panels and the port selection optimization of reconfigurable receive antenna arrays are often performed separately. Only the reflection phase, transmission phase, and amplitude parameters of the panel are adjusted individually, or only a subset of active ports of the receive array is optimized, failing to integrate panel configuration and port selection into a unified optimization framework. Since panel configuration directly affects the signal structure reaching the receiver, and port selection determines which signal components are ultimately utilized, there is a close coupling between the two. Optimizing one side alone inevitably results in a loss of gain available from the other, leaving significant room for a decrease in system information freshness.
[0005] Attempting to jointly optimize panel configuration and port selection presents significant challenges. The panel's reflection and transmission coefficients, as continuous variables, affect the channel gain, while the port activation vector is a discrete binary variable with cardinality constraints. Together, they determine the structure of the effective channel matrix, forming a highly coupled mixed-integer non-convex optimization problem. Such problems cannot be solved directly using standard convex optimization solvers and require specialized decomposition and iteration strategies.
[0006] To address the aforementioned coupling problem, a conventional approach is to use an alternating optimization method, decomposing the original problem into a panel configuration subproblem and a port selection subproblem, and solving them iteratively and alternately. However, in practice, it has been found that the coupling relationship between the two subproblems makes it difficult for the solutions of the alternating iterations to converge consistently. The objective functions of the subproblems are rewritten, making the alternating process prone to oscillations. More importantly, each time slot requires multiple calls to the convex optimization solver to process the subproblems, and the computation time increases dramatically with the number of users, array size, and number of time slots, making it difficult to meet the needs of online real-time decision-making.
[0007] To reduce online solution latency, existing technologies have proposed deep learning-based methods that attempt to directly output panel configuration or port selection strategies end-to-end via neural networks. However, existing networked methods iterate uniformly over all optimization variables, relying on standard continuously differentiable operators. They cannot directly handle the discrete characteristics of port selection, nor do they have the ability to characterize the recursive evolution of information freshness driven by random data arrival processes and rate thresholds. Furthermore, they fail to incorporate three different types of variables—panel configuration, port selection, and UAV trajectory planning—into a unified solution framework that can be executed with low latency during the inference phase.
[0008] Therefore, how to balance communication quality and information freshness while reducing solution overhead is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0009] To address the shortcomings of existing technologies, this invention provides an antenna-assisted low-latency air-to-space transmission method. By incorporating channel path reconstruction of the active reconfigurable transmission reflector panel, port reconstruction of the reconfigurable receiving antenna array, and UAV trajectory planning into a unified optimization framework, and employing a block-based iterative solution architecture that combines an outer second-order cone programming layer with an inner K-stage processing network, configuration parameters, port activation vectors, and UAV two-dimensional positions can be jointly updated. This ensures multi-user communication quality while reducing the average information freshness of the system and shortening the online solution time.
[0010] To achieve the above objectives, the present invention adopts the following technical solution:
[0011] An antenna-assisted low-latency air-to-space transmission method includes the following steps:
[0012] Step 1: Obtain the initial system parameters and construct a data transmission system model based on the initial system parameters. The initial system parameters include at least the UAV's flight altitude, the UAV's flight area, the array size of the active reconfigurable transmission reflector, the transmit power budget of the active reconfigurable transmission reflector, the number of candidate ports of the reconfigurable receiving antenna array, the number of active ports of the reconfigurable receiving antenna array, the number of ground users, the rate threshold, and channel statistics parameters.
[0013] Step 2: Construct the direct channel from the ground user to the base station, the reflection channel from the ground user to the active reconfigurable transmission reflector panel, the cascaded reflection channel from the active reconfigurable transmission reflector panel to the base station, the transmission channel from the active reconfigurable transmission reflector panel to the airborne receiver, and the port spatial correlation matrix of the reconfigurable receiving antenna array; based on non-orthogonal multiple access uplink reception, perform serial interference cancellation and minimum mean square error merging on the base station path to obtain the base station path rate; wherein the base station path includes the direct path of the ground user's transmitted signal to the base station via the direct channel, and the reflection path of the ground user's transmitted signal to the base station via the reflection channel and the cascaded reflection channel, and the signals of the direct path and the reflection path are superimposed at the base station; perform serial interference cancellation on the airborne receiver path to obtain the airborne receiver path rate; wherein the airborne receiver path is the path of the ground user's transmitted signal to the airborne receiver via the reflection channel and the transmission channel; recursively deduce the user information freshness based on the base station path rate, the airborne receiver path rate, the data packet arrival indication, the packet availability indication, and the rate threshold, and calculate the system average information freshness;
[0014] Step 3: Using the configuration parameters of the active reconfigurable transmission reflector panel, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV as optimization variables, and minimizing the average information freshness of the system as the objective, a mixed integer nonconvex joint optimization problem is constructed under the constraints of parameter value range, UAV flight constraints, port selection constraints, and transmit power budget constraints of the active reconfigurable transmission reflector panel.
[0015] Step 4: Based on the K-stage processing network, perform online block-based iterative solution in the outer and inner layers, and output the updated configuration parameters of the active reconfigurable transmission reflector panel, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV. The outer layer fixes the configuration parameters and port activation vector and jointly updates the two-dimensional position of the UAV in each time slot; the inner layer fixes the two-dimensional position of the UAV and updates the configuration parameters and port activation vector by time slot by the K-stage processing network.
[0016] Step 5: For each time slot, calculate the base station path rate and airborne receiver path rate based on the updated configuration parameters, port activation vector, and UAV 2D position output in Step 4. Determine whether the task is successful based on the calculated base station path rate and airborne receiver path rate, and update the user information freshness. After all time slots are completed, output the system average information freshness and the corresponding transmission strategy.
[0017] Furthermore, in step 1, the channel statistical parameters include the thermal noise power of the base station receiver, the thermal noise power of the airborne receiver, the noise figure of the power amplifier per component, the Rice factor, and the path loss index; the system initial parameters also include the total number of UAV time slots, the energy splitting parameter, and the boundary parameters for the amplitude utilization factor; the data transmission system model is established in a unified coordinate system.
[0018] Furthermore, in step 3, the configuration parameters include reflection amplitude, transmission amplitude, energy splitting parameter, amplitude utilization factor, reflection phase, and transmission phase;
[0019] The reflection amplitude and reflection phase constitute the reflection configuration matrix:
[0020]
[0021] The transmission amplitude and transmission phase constitute the transmission configuration matrix:
[0022]
[0023] In the formula, For active reconfigurable transmission reflective panels in time slots The reflection configuration matrix, For diagonal matrix operations, For the first Each component in the time slot The amplitude of reflection, , For the first Each component in the time slot The reflection phase, For active reconfigurable transmission reflective panels in time slots The transmission configuration matrix, For the first Each component in the time slot The transmission amplitude, , For the first Each component in the time slot The transmission phase, , For the array size of actively reconfigurable transmission reflective panels, The imaginary unit; For the first Each component in the time slot The energy splitting parameters, For the first Each component in the time slot Amplitude utilization factor, The upper limit of hardware amplification per component, These are the boundary parameters for the energy splitting parameter and the amplitude utilization factor.
[0024] Furthermore, in step 2, the signals from the direct path and the reflected path are superimposed at the base station and then received by the reconfigurable receiving antenna array. The reconfigurable receiving antenna array selects the effective channel of the base station path corresponding to the active port from the full-dimensional channel of the base station through a selection matrix.
[0025] When the base station path performs serial interference cancellation and minimum mean square error merging, the interference plus noise covariance matrix is composed of the base station receiver thermal noise covariance and color noise covariance. The color noise covariance is formed by the power amplifier noise of each element. The power amplifier noise of each element is modulated by the reflection configuration matrix and then propagated to the base station through the cascaded reflection channel.
[0026] Furthermore, in step 4, the process of updating the two-dimensional position of the drone in the outer layer includes:
[0027] Performing a first-order Taylor expansion on the path loss term yields a lower-bound affine function for the UAV's two-dimensional position:
[0028]
[0029]
[0030] In the formula, In time slot No. The distance from a ground user to an active reconfigurable transmission reflector panel. In time slot The distance from the active reconfigurable transmission reflector panel to the base station, For the outer iteration round, For the first The first iteration of the outer layer The distance from a ground user to an active reconfigurable transmission reflector panel. For the first The distance from the active reconfigurable transmission reflector panel to the base station during the outer layer iteration process. For the first The slack variable of the distance from a ground user to the drone. Let the distance from the drone to the base station be a slack variable. The path loss index;
[0031] Introduction and As a second-order cone constraint:
[0032]
[0033]
[0034] In the formula, For the first The location of each ground user. For drones in time slots Two-dimensional position, The altitude at which the drone flies. Location of the base station;
[0035] A weighted objective function is constructed using the square of the user information freshness as a weighting coefficient. The upper limit constraint of UAV displacement and the UAV flight area constraint are added to form a second-order cone programming subproblem and solve it to obtain the UAV's two-dimensional position after the outer layer iterative update.
[0036] Furthermore, it also includes training the K-stage processing network, constructing a supervision signal based on the task success flag, packet arrival indication, packet availability indication, and user information freshness, using the supervision signal as loss to train the learnable parameters of each stage, and freezing the network parameters after training is completed, wherein the task success flag is generated based on the base station path rate, airborne receiver path rate, and rate threshold.
[0037] The training objective function of the K-stage processing network adopts a hierarchical weighted sum form. For each stage, the smooth decision result of the output is weighted by the square of the freshness of user information and accumulated to obtain the loss of each stage. The loss of each stage is summed and normalized after applying increasing weights according to the stage number to obtain the total loss.
[0038] Furthermore, the K-stage processing network comprises K sequentially executed stages, with the output of each stage serving as the input to the next stage, and the output of the final stage serving as the inner layer update result. The input to the first stage includes the initial port score, the UAV's two-dimensional position obtained from the outer layer update, the configuration parameters obtained from the previous inner layer update, the port activation vector obtained from the previous inner layer update, and the channel state information of the current time slot. The update process of each stage sequentially includes a reflection phase and transmission phase alignment module, a stepwise energy splitting parameter and amplitude utilization factor projection gradient update module, a cross-block attention modulation module, and a graph attention and differentiable port selection module.
[0039] Under the condition of fixing the reflection amplitude and transmission amplitude in the current stage input, the reflection phase and transmission phase alignment module constructs auxiliary variables for each ground user, constructs a service quality weighting with a learnable slope as a parameter, multiplies the service quality weighting with the auxiliary variables to obtain the phase alignment coefficient, and updates the reflection phase and transmission phase of each element according to the phase alignment coefficient.
[0040] The stepwise energy splitting parameter and amplitude utilization factor projection gradient update module calculates the partial derivatives of the loss function with respect to the energy splitting parameter and amplitude utilization factor under the condition of fixed reflection phase and transmission phase. It calculates the pre-projection update values of the energy splitting parameter and amplitude utilization factor according to the learnable step size and the partial derivatives. The pre-projection update values of the energy splitting parameter and amplitude utilization factor are projected to their respective value ranges by the box projection operator to obtain the updated energy splitting parameter and amplitude utilization factor.
[0041] The cross-block attention modulation module constructs panel state feature vectors, trajectory feature vectors, and port feature vectors. The three types of feature vectors are mapped to a shared latent space through independent linear projection and layer normalization, respectively, resulting in three tokens. The three tokens form a token matrix. Multi-head attention operation is applied to the token matrix. The output of the multi-head attention operation is linearly transformed, hyperbolic tangent activated, and multiplied by a preset upper limit to obtain the modulation bias and port score bias of the active reconfigurable transmission reflection panel. The modulation bias is fed back to the element-wise energy splitting parameter and amplitude utilization factor projection gradient update module. The port score bias is superimposed on the graph attention port score in the next stage.
[0042] The graph attention and differentiable port selection module is used to construct an undirected graph containing port nodes, user nodes, and global context nodes. Each node is assigned an initial embedding, which is iteratively updated in a two-layer graph attention network. Port node embeddings are selected from the updated node embeddings, and a multilayer perceptron is used to obtain the score for each port. After superimposing the port score bias from the previous stage, a Gumbel perturbation is introduced to calculate soft selection, ensuring that the sum of the soft selection values equals... Select the highest-rated port from all ports according to the hard selection rule. The set of active ports consists of several ports. To determine the number of activated ports, a pass-through gradient estimation method is used, enabling the gradient corresponding to hard selection to propagate back through the soft selection path.
[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0044] (1) This invention integrates the channel path reconstruction of the active reconfigurable transmission reflective panel, the port reconstruction of the reconfigurable receiving antenna array, and the two-dimensional position planning of the UAV into a unified optimization framework, and performs joint solution with the system average information freshness as the optimization objective. The configuration parameters of the panel determine the amplitude and phase distribution of the signal after propagation through the reflection link and the transmission link. The port selection determines which spatial locations the base station extracts signal components from. The UAV position simultaneously changes the distance and channel conditions of the two propagation links from the ground user to the panel and from the panel to the base station. The three types of variables act on different stages of channel propagation and have a close coupling relationship. Optimizing any one type of variable alone can only improve the local stage of channel propagation and cannot make the gains of the three types of variables match each other. This invention, through joint optimization, enables the panel's channel path reconstruction, the port's spatial diversity extraction, and the UAV's adjustment of the propagation distance to work synergistically under the same optimization objective, avoiding the gain loss caused by unilateral optimization, thereby reducing the system's average information freshness as a whole.
[0045] (2) This invention replaces the non-differentiable task success flag with a smooth approximation function based on a sigmoid function, and constructs a supervision signal based on the task success flag, packet arrival indication, packet availability indication, and user information freshness. The K-stage processing network is trained using the supervision signal as the loss. The recursion of information freshness is driven by the task success flag, which is determined by whether the dual-path rate has reached a threshold. The rate threshold decision is essentially a non-differentiable step function. The step function cannot directly participate in backpropagation, which means that the K-stage processing network cannot be trained by gradient descent. This invention uses a sigmoid function to smoothly approximate the rate threshold decision, establishing a continuous and differentiable mapping relationship between the dual-path rate and the task success probability. At the same time, the square of information freshness is used as a weighting coefficient, so that users with higher information freshness have a greater weight in the loss function. Users with higher information freshness have longer packet waiting times and are more eager to improve the marginal probability of task success. The square weighting makes the network prioritize allocating configuration resources to users with the greatest information freshness pressure during training, thereby achieving a faster decrease in the system average information freshness.
[0046] (3) This invention uses a K-stage processing network to sequentially perform phase-weighted alignment, gradient update of the element-wise energy splitting parameter and amplitude utilization factor, cross-block attention modulation, graph attention scoring, and differentiable port selection, thereby achieving joint updates of configuration parameters and port activation vectors. Phase-weighted alignment determines the reflection and transmission phases of each element based on the phase relationship of the equivalent channel, so that the signal is coherently superimposed at the target receiver after being modulated by the panel; gradient update of the element-wise energy splitting parameter and amplitude utilization factor adjusts the energy allocation ratio and amplitude utilization degree of each element according to the gradient of the loss function, so that the amplification power of the panel is concentrated on the element that contributes the most to the freshness of information; cross-block attention modulation maps the panel state, trajectory state, and port state to the shared latent space through multi-head attention operation, so that the coupling relationship between the panel configuration branch and the port selection branch is explicitly modeled; graph attention scoring constructs the correlation relationship between ports through the port space correlation matrix, and outputs the set of activated ports through differentiable port selection. The four sub-processes are executed sequentially according to their functions. The output of the previous sub-process serves as the input of the next sub-process, allowing the panel configuration parameters and port selection to be updated alternately and adapted to each other within the same stage. This avoids oscillations caused by inconsistencies in sub-problems during alternating optimization, thereby accelerating the convergence speed and reducing the solution overhead.
[0047] (4) This invention uses second-order cone programming to update the two-dimensional position of the UAV in the outer layer and K-stage processing network to update the panel configuration parameters and port activation vectors in the inner layer. Block iterative solution is achieved through the alternating execution of the outer and inner layers. The change of the UAV's two-dimensional position affects the distances of the two links from the user to the panel and from the panel to the base station. The negative path loss exponent term of the distance makes the path loss and UAV position present a non-convex relationship. This invention performs a first-order Taylor expansion on the negative path loss exponent term of the distance to obtain a lower bound affine function for the UAV position, and introduces a distance relaxation variable to transform the non-convex constraint into a second-order cone constraint, thereby transforming the UAV trajectory subproblem into a standard second-order cone programming problem, which can be efficiently solved by existing solution tools and satisfies flight constraints. The panel configuration parameters and port activation vectors are both high-dimensional variables and the port activation vectors are discrete variables. This invention entrusts the joint update of the two to the K-stage processing network. The network learns the coupling relationship between variables through offline training. During online inference, it only needs a fixed number of forward calculations to output the entire transmission strategy, without the need to call the convex optimization solver online. The outer second-order cone programming ensures the feasibility and optimality of the UAV trajectory, while the inner K-stage processing network ensures the consistency and low latency of configuration parameters and port selection. The two are executed alternately, which reduces the average information freshness of the system while ensuring the quality of multi-user communication and shortens the solution time. Attached Figure Description
[0048] Figure 1 A flowchart of the antenna-assisted low-latency air-to-space transmission method provided by the present invention;
[0049] Figure 2 This is a comparison graph showing the changes in the average information freshness of the system with the rate threshold when using the method of the present invention and the comparative method, respectively.
[0050] Figure 3 This diagram shows the scheduling results of successful transmission by ground users in each time slot using the method of the present invention via the base station path and the airborne receiver path in each time slot.
[0051] Figure 4 The diagram shows the scheduling results of successful transmission by ground users in each time slot using Method 1 via the base station path and the airborne receiver path.
[0052] Figure 5 The diagram shows the scheduling results of successful transmission by ground users in each time slot via the base station path and the airborne receiver path under Method 3;
[0053] Figure 6 The diagram shows the scheduling results of successful transmission by ground users in each time slot via the base station path and the airborne receiver path under Method 6. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] Example
[0056] refer to Figure 1 This embodiment provides an antenna-assisted low-latency air-to-space transmission method, including the following steps:
[0057] Step 1: Obtain initial system parameters. In this embodiment, the initial system parameters include: the UAV's flight altitude. The flight area of drones (The range of two-dimensional flyable space constrained by latitude and longitude boundaries), and the array size of the active reconfigurable transmission reflective panel (hereinafter referred to as "panel"). Panel transmit power budget Hardware amplification limit per component (Maximum signal amplitude that a single-component power amplifier can provide), number of candidate ports for a reconfigurable receiver antenna array Number of active ports of the reconfigurable receiving antenna array Number of ground users Rate threshold Total number of time slots for drones Boundary parameters for energy splitting parameters and amplitude utilization factors And channel statistical parameters. These channel statistical parameters include the base station receiver thermal noise power. Airborne receiver thermal noise power Noise figure per component of power amplifier Rice factor and path loss index .in, For the panel edge Number of components in the axial direction For the panel edge Number of components in the axial direction For reconfigurable receiver antenna array along Number of candidate ports in the axial direction. For reconfigurable receiver antenna array along Number of candidate ports in the axial direction. In this embodiment, the preset minimum positive number is used. =0.01, used to prevent a transmission path from being completely shut down. The values of each parameter can be flexibly configured according to the actual system deployment scenario.
[0058] Based on the obtained initial system parameters, a data transmission system model is constructed in a unified coordinate system. Let the set of ground users be denoted as... , No. The location of each ground user is denoted as The probability of a data packet arriving (the probability of generating a new data packet per time slot according to a Bernoulli process) is denoted as... The data transmission system includes a rotary-wing drone, a panel mounted beneath the drone, a base station configured with a reconfigurable receiving antenna array, and... There are several ground users. The UAV side is equipped with an airborne edge computing unit and an airborne receiver, while the base station side is equipped with a base station edge computing unit. The airborne edge computing unit performs serial interference cancellation decoding on the signal received by the airborne receiver. The reconfigurable receiving antenna array outputs the received signal to the base station receiver, where the base station edge computing unit performs channel estimation, serial interference cancellation, and minimum mean square error (MME) combining decoding on the signal received by the base station receiver. The airborne receiver selects the signal with the strongest channel quality in each time slot. Each ground user performs serial interference cancellation decoding. The decoding capacity of the airborne edge computing unit is less than The two-dimensional position of the drone is denoted as... The location of the base station is recorded as The two-dimensional aperture size of the reconfigurable receiving antenna array is , For along Normalized aperture length in the axial direction, For along Normalized aperture length in the axial direction, The wavelength is the signal wavelength.
[0059] Within each time slot, the reflection configuration matrix of the panel is as follows:
[0060]
[0061] The transmission configuration matrix is as follows:
[0062]
[0063] in, For the panel in time slot The reflection configuration matrix, For diagonal matrix operations, For the first Each component in the time slot The amplitude of reflection, For the first Each component in the time slot The reflection phase, For the panel in time slot The transmission configuration matrix, For the m-th element in time slot The transmission amplitude, For the first Each component in the time slot The transmission phase, , It is the imaginary unit.
[0064] No. The reflection amplitude of each element is according to Perform parameterization, transmission amplitude according to Parameterization is performed, where, For the first Each component in the time slot The energy splitting parameters, For the first Each component in the time slot Amplitude utilization factor, This represents the upper limit of hardware amplification per component.
[0065] Let the first The thermal noise injected by the power amplifier on each component is Its mean is variance is The complex Gaussian distribution, i.e.:
[0066]
[0067] In the formula, The variance of the thermal noise injected into each power amplifier is calculated as follows:
[0068]
[0069] In the formula, The noise figure of the power amplifier per component, Boltzmann's constant, For reference temperature, this embodiment uses... , This refers to the system bandwidth.
[0070] Arrange the thermal noise of the power amplifier with all components as a vector. Synthetic noise in the direction of reflection according to Synthetic noise propagating from the panel to the base station via a cascaded reflection channel in the transmission direction according to It propagates from the panel to the drone via the transmission channel.
[0071] Step 2: Construct the direct channel from the ground user to the base station, the reflection channel from the ground user to the active reconfigurable transmission reflector panel, the cascaded reflection channel from the active reconfigurable transmission reflector panel to the base station, the transmission channel from the active reconfigurable transmission reflector panel to the airborne receiver, and the port spatial correlation matrix of the reconfigurable receiving antenna array; based on non-orthogonal multiple access uplink reception, perform serial interference cancellation and minimum mean square error merging on the base station path to obtain the base station path rate; wherein the base station path includes the direct path of the ground user's transmitted signal to the base station via the direct channel, and the reflection path of the ground user's transmitted signal to the base station via the reflection channel and the cascaded reflection channel, and the signals of the direct path and the reflection path are superimposed at the base station; perform serial interference cancellation on the airborne receiver path to obtain the airborne receiver path rate; wherein the airborne receiver path is the path of the ground user's transmitted signal to the airborne receiver via the reflection channel and the transmission channel; recursively deduce the user information freshness based on the base station path rate, the airborne receiver path rate, the data packet arrival indication, the packet availability indication, and the rate threshold, and calculate the system average information freshness.
[0072] No. Direct channels from ground users to base stations Represented as:
[0073]
[0074] No. Reflection channel from ground user to panel Represented as:
[0075]
[0076] Cascaded Reflection Transmission Channel from Panel to Base Station Represented as:
[0077]
[0078] In the formula, For reference path gain, For the first The distance from each ground user to the base station, In time slot No. The distance from each ground user to the panel, In time slot The distance from the panel to the base station, This is the path loss index. For direct channel in time slot Rice's small-scale fading component, For the reflection channel in time slot Rice's small-scale fading component, For cascaded reflection channels in time slots Rice's small-scale fading component, for 3D complex vector space, for 3D complex vector space, for OK The complex matrix space of columns.
[0079] The transmission channel from the panel to the airborne receiver adopts a near-field spherical wave model, and the distance from each component to the airborne receiver is calculated separately. Each component in the time slot The distance to the airborne receiver is denoted as . No. The path loss coefficient of the transmission channel corresponding to each element is Phase factor is ,in The imaginary unit is used to represent the number. The path loss coefficients of each component are arranged in a column vector. ,Will The phase factors of each element are arranged as a column vector. Transmission channel from panel to airborne receiver Represented as:
[0080]
[0081] In the formula, For Hadama accumulation.
[0082] Port spatial correlation matrix of reconfigurable receiving antenna array Calculated according to the Jakes model The Middle Line number Column elements Represented as:
[0083]
[0084] In the formula, It is a zero-order Bessel function of the first kind. For the first Two-dimensional physical location of each candidate port For the first Two-dimensional physical location of each candidate port Port space correlation matrix Perform square root decomposition, multiply the Ricean fading non-line-of-sight component with the square root decomposition result, and obtain the equivalent receive channel with port correlation.
[0085] No. When the transmitted signal from a ground user reaches the base station, it is a superimposed signal obtained by adding the signal from the direct path and the signal from the reflected path. The signal from the reflected path is composed of the signal from the first ground user. The transmitted signal from a ground user reaches the panel via a reflection channel, is modulated by the panel's reflection configuration matrix, and then reaches the base station via a cascaded reflection channel. The equivalent channel corresponding to each candidate port for:
[0086]
[0087] In the formula, For the first The number of ground users to the first Direct channels for candidate ports for The conjugate transpose of . For the panel up to the first Cascaded reflection channels of candidate ports For the first A ground user's reflection channel to the panel.
[0088] Time slots The port activation vector of the reconfigurable receiving antenna array is , for The The element represents the element. One candidate port in the time slot The activation status. The set of activated ports is... , No. The transmitted signals from each ground user are superimposed after reaching the base station via direct and reflected paths. This is achieved by selecting a matrix. Select the effective channel corresponding to the active port from the base station's full-dimensional channel. for:
[0089]
[0090] In the formula for The transpose of .
[0091] At the airborne receiver, because the panel is horizontally mounted under the drone fuselage, it obstructs the direct transmission link from the user to the airborne receiver; only a transmissive transmission link is retained. The transmitted signal from a ground user reaches the panel via a reflection channel, is modulated by the panel's transmission configuration matrix, and then reaches the airborne receiver via a transmission channel. The corresponding complete equivalent channel... for:
[0092]
[0093] In the formula, Transmission channel from panel to airborne receiver The conjugate transpose of .
[0094] Serial interference cancellation and minimum mean square error merging are performed on the base station path. The decoding order of serial interference cancellation on the base station side is denoted as follows: According to the effective channels of ground users in various regions norm Determined by sorting from largest to smallest. The combined signal-to-interference-plus-noise ratio of each decoding user for:
[0095]
[0096] Corresponding base station path rate for:
[0097]
[0098] In the formula, For the first Transmission power of each ground user For the first Effective channels for each ground user for The conjugate transpose of . for The inverse matrix, For the first The interference plus noise covariance matrix of each decoding user is expressed as:
[0099]
[0100]
[0101] in, This refers to the thermal noise power of the base station receiver. for The identity matrix, The color noise covariance matrix is... for The conjugate transpose of . for The conjugate transpose of .
[0102] Perform serial interference cancellation on the airborne receiver path. Let the decoding sequence for serial interference cancellation on the airborne receiver side be denoted as follows: Based on the complete equivalent channel from ground users to airborne receivers in various locations modulus The order is determined from largest to smallest, and the airborne receiver only decodes the first few items scheduled for decoding. The ground users are sorted, the first Each decoder user signal-to-interference-plus-noise ratio for:
[0103]
[0104] Corresponding airborne receiver path rate for:
[0105]
[0106] In the formula, The combined noise power is the noise from the power amplifiers of each component on the panel, modulated by the transmission configuration matrix, and then propagated through the transmission channel to the airborne receiver. This refers to the thermal noise power of the airborne receiver.
[0107] Each user maintains at most one data packet to be transmitted, and a buffer is maintained according to a preemptive last-come-first-served rule. (In time slots) Initially based on packet arrival probability A new arriving data packet is generated, overwriting any untransmitted old packets; the arrival indication is denoted as... The package can be indicated as The buffer waiting time is The success sign of the task is Package available instructions Update iteratively according to the following rules:
[0108]
[0109] Buffer wait time Update iteratively according to the following rules:
[0110]
[0111] When either the base station path rate or the drone path rate reaches the rate threshold and there are data packets in the buffer, the mission is successful. .
[0112] No. Individual users in time slots Information freshness It evolves recursively according to the following rules:
[0113]
[0114] This sets the upper limit for the freshness of user information.
[0115] System average information freshness according to Perform the calculation.
[0116] Step 3: Using the panel configuration parameters, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV as optimization variables, and minimizing the average information freshness of the system as the objective, a mixed integer nonconvex joint optimization problem is constructed under the constraints of parameter value range, UAV flight constraints, port selection constraints, and panel transmit power budget constraints.
[0117] Let the set of panel configuration parameter variables be... :
[0118]
[0119] The set of drone trajectory variables is :
[0120]
[0121] The set of port activation vectors for a reconfigurable receiving antenna array is :
[0122]
[0123] Let the set of all feasible transmission strategies be denoted as . .
[0124] The mixed-integer nonconvex joint optimization problem is represented as:
[0125]
[0126] The constraints are:
[0127]
[0128]
[0129]
[0130]
[0131]
[0132]
[0133]
[0134]
[0135]
[0136] In the formula, For the panel number The input power of each component; For the first A ground user in a time slot The transmission power, In time slot , No. The number of ground users to the first The reflection channel coefficient of each panel element, The maximum flight speed of the drone, The duration is for a single time slot; Designated area for drone flights; Budget for the panel's transmit power; It is a preset minimum positive number; For feasible transmission strategies, Denotes the random variable representing the Bernoulli arrival process. Indicates channel implementation, To implement feasible transmission strategies Arrival process and channel implementation The average information freshness of the system under the following conditions It represents the mathematical expectation.
[0137] Step 4: Based on the K-stage processing network, perform online block-based iterative solution in the outer and inner layers, and output the updated configuration parameters of the active reconfigurable transmission reflector panel, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV. The outer layer fixes the configuration parameters and port activation vector and jointly updates the two-dimensional position of the UAV in each time slot; the inner layer fixes the two-dimensional position of the UAV and updates the configuration parameters and port activation vector by time slot by the K-stage processing network.
[0138] The specific process of the outer iterative solution is as follows:
[0139] Replace the non-differentiable task success flag with a smooth approximation function, denoted as the smooth approximation task success flag. :
[0140]
[0141]
[0142] in, It is a differentiable sigmoid function. A value greater than 0 represents an adjustable sharpness coefficient, used to adjust the smoothness of the S-shaped function. For drone dispatch instructions, when the first Individual users in time slots When the airborne receiver performs serial interference cancellation decoding, =1, otherwise Because the airborne receiver only processes the signal in the time slot with the strongest channel quality, Each ground user performs decoding, therefore satisfying . For the first Individual users in time slots The smoothing probability that at least one of the base station path rate and the airborne receiver path rate reaches a threshold when transmitting signals. For the first Individual users in time slots Base station path rate, For the first Individual users in time slots The path rate of the airborne receiver.
[0143] Task success flag based on smooth approximation Construct a task success estimation function, and then use this function to construct a second-order cone programming subproblem:
[0144] Let the outer iteration round be . At the start of the current iteration, the input is known. , , , . For the drones obtained in the previous outer layer iteration in the time slot Two-dimensional position, This is the reflection configuration matrix obtained from the previous inner iteration. This is the transmission configuration matrix obtained from the previous inner layer iteration. This is the port activation vector obtained from the previous inner layer iteration.
[0145] No. When iterating through the outer layer of the wheel, the first... The distance from each ground user to the panel is The distance from the panel to the base station is :right and A first-order Taylor expansion yields a lower-bound affine function for the UAV's position:
[0146]
[0147]
[0148] Distance slack variable from ground user to drone The slack variable of distance from the drone to the base station ,satisfy:
[0149]
[0150]
[0151] by To construct a weighted objective function for the weighted coefficients, and to add upper limit constraints on displacement and flight region constraints, we obtain the following second-order cone programming subproblem:
[0152]
[0153] The constraints are:
[0154]
[0155]
[0156]
[0157]
[0158] The task success estimation function is obtained based on the first-order Taylor expansion, representing the function in the first order. In the outer layer iteration, slack variables are used. and The approximate value of the success probability of the task is denoted by .
[0159] Solve this second-order cone programming subproblem using second-order cone programming solvers such as Gurobi to obtain the outermost... Two-dimensional position update value of the wheeled drone ,Will As a read-only context, it is fed into the inner K-stage processing network, where The gradient is truncated during backpropagation.
[0160] The inner layer is solved iteratively through a K-stage processing network. The K-stage processing network updates the panel configuration and port activation vectors. The K-stage processing network consists of K stages executed sequentially. The output of the previous stage is used as the input of the next stage, and the output of the last stage is used as the inner layer update result.
[0161] The inputs for Phase 1 include: initial port score values and values obtained from the outer update. The results of the previous inner layer update , , and the channel state information of the current time slot, including the reflection channel. Cascaded Reflection Channel Transmission Channel and direct channel In this embodiment, the initial value of the port score is 0. The temperature coefficient of the stage is denoted as Parameterization is performed using the product of the learnable scalar and the global temperature coefficient. For simplicity, the time slot index is omitted from the parameters during the inner layer update process. All variables represent the current time slot. The value of .
[0162] The update process of each stage in the K-stage processing network includes four modules in sequence: reflection phase and transmission phase alignment module, element-wise energy splitting parameter and amplitude projection gradient update module, cross-block attention modulation module, graph attention and differentiable port selection module.
[0163] The reflection phase and transmission phase alignment module is used to align the reflection phase with the reflection amplitude output from the previous stage. Transmission amplitude The channel state of each user updates the reflection and transmission phases of each element, and outputs the updated reflection phase. With transmission phase The specific process is as follows:
[0164] With a fixed reflection amplitude Transmission amplitude Under these conditions, construct auxiliary variables for each user. Calculated according to the following rules:
[0165]
[0166] in, and These are the signal power term and interference power term from the previous stage, respectively. This is a scalar approximation of the equivalent noise power of the component power amplifier after propagation to the base station via a cascaded reflection channel. This is the reflection configuration matrix output from the previous stage.
[0167] by Learnable slope of the stage Build a service quality weighted system for parameters :
[0168]
[0169] In the formula, For the first The path rate of each user in the previous stage (take the larger of the base station path rate and the airborne receiver path rate). This is the rate threshold.
[0170] Phase alignment coefficient for:
[0171]
[0172] No. The reflection phase of each element is updated according to the weighted alignment rule as follows:
[0173]
[0174] In the formula, This represents the argument operation for complex numbers. This is a conjugate operation. , For the first stage The signal transmitted by the ground user is transmitted through the panel. After each component is reflected, it then passes through the corresponding set of activation ports. The equivalent channel response of all active ports to the base station. for The Middle Line number Column elements, For reflection channel The The element represents the element. The number of ground users to the panel The reflection channel coefficients of each element.
[0175] The transmission phase is updated using the same weighted alignment rule as the reflection phase, only replacing the base station path-dependent channels with the UAV path-dependent channels. To stabilize training, the updated transmission phase is mixed with the transmission phase from the previous stage using a momentum coefficient. In this embodiment, the momentum coefficient is set to 0.9. The mixed result is used as the final transmission phase output for this stage.
[0176] The element-wise energy splitting parameter and amplitude utilization factor projection gradient update module is used to update the energy splitting parameters and amplitude utilization factors of each element under the condition of fixed phase. The input of this module is the energy splitting parameters of the previous stage. and amplitude utilization factor The output is the updated energy splitting parameters. and amplitude utilization factor The specific process is as follows:
[0177] according to and Parameterize the reflection and transmission amplitudes at a fixed phase. , Under the conditions and Update according to the projection gradient rule respectively:
[0178]
[0179]
[0180] In the formula, These are the energy splitting parameters from the previous stage. This is the amplitude utilization factor of the previous stage. For the first Pre-projection update values of stage energy splitting parameters. For the first The stage amplitude utilizes the pre-projection update value of the factor. For the first The learnable step size of the stage energy splitting parameters; For the first The learnable step size of the stage amplitude utilization factor; For the first The loss function for each stage For loss function about The partial derivatives, For loss function about The partial derivatives of the two types are calculated through backpropagation. The bias is used to modulate the energy splitting parameters fed back by the cross-block attention modulation module. The amplitude of the feedback from the cross-block attention modulation module is biased using a factor modulation method.
[0181] Finally, update the value. and Project them separately into their respective feasible regions:
[0182]
[0183]
[0184] in, For box projection operators, Indicates will Projected to interval Inside, that is, when Time to take ,when Time to take Otherwise take itself; Indicates will Projected to interval Inside, that is, when Time to take ,when Take 1 if it is true, otherwise take 1. itself.
[0185] The cross-block attention modulation module extracts multimodal features from panel state, trajectory state, and port state, and generates modulation biases through a multi-head attention mechanism. The module's inputs are the current stage's configuration state and trajectory context, and its outputs are the modulation biases fed back to the projection gradient update module and the biases superimposed on the port scores. The specific process is as follows:
[0186] Construct panel state feature vectors, trajectory feature vectors, and port feature vectors as inputs.
[0187] Panel state feature vector:
[0188]
[0189] In the formula, For the first The vector formed by the reflection amplitudes of each element in the stage. For the first The vector formed by the transmission amplitudes of each element in the stage. For the first The vector formed by the reflection phases of each element in the stage. For the first The vector formed by the transmission phases of each element in the stage.
[0190] The trajectory feature vector is:
[0191]
[0192] The port feature vector is:
[0193]
[0194] In the formula, This is the output of the pass-through gradient estimation from the previous stage.
[0195] The three types of feature vectors are mapped to the shared latent space through independent linear projection and layer normalization, respectively. Stacked into a token matrix In this system, independent linear projection is a fully connected layer, layer normalization is a layer standardization operation, and the dimension of the shared latent space is 64. Multi-head attention is applied to the token matrix, and the output undergoes a linear transformation and hyperbolic tangent activation, multiplied by a preset upper limit. The panel modulation bias is obtained. and port scoring bias :
[0196]
[0197] in, The modulation bias is fed back to the energy splitting parameters and the amplitude utilization factor projection gradient update module. The bias vector for the energy splitting parameter modulation, its first... element For the Update the energy splitting parameters of each component; The amplitude is modulated by the factor, and its first... element For the The amplitude of each component is updated using the factor, and the port scoring bias is applied. This is then superimposed onto the graph attention port scoring in the next stage. The trajectory feature vectors undergo gradient truncation, providing only read-only context. For the first The learnable linear transformation layer of the stage maps the multi-head attention output to the modulation bias space. The hyperbolic tangent activation function is used. This is a multi-head attention operation.
[0198] The graph attention and differentiable port selection module generates scores for each port based on port space correlation and global context, and outputs the set of activated ports using a differentiable pre-K selection operator. The module's inputs are the port space correlation matrix, user state, and global context; the output is a hard selection indicator. and through gradient estimation output The specific process is as follows:
[0199] Constructing an undirected graph:
[0200]
[0201] in For a set of port nodes, For a set of user nodes, For global context nodes; edges between port pairs Carrying port space-related matrix elements; edges from users to port nodes. Carry information freshness weights and channel requirement differences; edges from global context nodes to all port nodes. Broadcast panel status and track status.
[0202] Node embeddings are iteratively updated in the two-layer graph attention network according to the following rules:
[0203]
[0204] in, For the first Layer nodes Embedded vector, This is the index of the iteration layer of the graph attention network. ; For the activation function, a linear unit with leakage correction is used in this embodiment. No. Layer nodes With nodes The attention weights between elements are calculated using a graph attention mechanism combined with edge weight scaling and leakage-corrected linear unit activation. For the first The learnable weight matrix of the layer; For the first Layer nodes Embedded vector, This is the initial embedding of the node.
[0205] After two layers of graph attention network, the first... Rating of each port for:
[0206]
[0207] In the formula, For a multilayer perceptron, the node embedding of the port is mapped to a scalar score; For the first Phase 1 The node embedding vectors of each port are obtained by iteratively updating the formula through a two-layer graph attention network; For the first Port scoring bias in the phase The Each element is output by the cross-block attention modulation module.
[0208] Introducing the Gumbel perturbation:
[0209]
[0210] In the formula, For the first Gumbel perturbation on each port For the first Uniformly distributed random variables generated from each port.
[0211] Calculate soft selection:
[0212]
[0213] In the formula, For the first The soft selection value for each port. For the first The rating of each port, For the first Temperature coefficient of the stage.
[0214] Cardinality matching for soft selection:
[0215]
[0216] This is the soft selection vector after cardinality matching. This is a soft selection vector.
[0217] Press hard front Select the rule and choose the highest score from the scores of each port. Each port is used as an active port to obtain the set of active ports. :
[0218]
[0219] in, This indicates selecting the largest value from the input vector. The index of each element For port rating vectors, This is the Gumbel perturbation vector.
[0220] Hard selection instruction Defined as , Hard selection indicator vector The The element, when the first element The value is 1 when a port is selected, and 0 otherwise.
[0221] Final output by Implement pass-through gradient estimation so that the gradient corresponding to hard selection propagates back through the soft selection path.
[0222] in, To stop gradient calculation, i.e., to take the input value during forward propagation and not calculate the gradient during backward propagation, this operation prevents gradient calculation during forward propagation. This is equivalent to a hard selection indicator; during backpropagation, the gradient passes through a soft selection indicator. This allows the gradient corresponding to hard selection to propagate backward through the soft selection path.
[0223] At the output of each stage in the inner layer, box projection is performed on the amplitude utilization factor and energy splitting parameter according to their value ranges; single-mode constraints are performed on the reflection phase and transmission phase by angle-taking operations; and hard fronting is applied to the port activation vector. Choose from A fixed selection from the candidates One activation port.
[0224] For the time-slot-by-time transmit power budget constraint of the active reconfigurable transmit reflector panel, a post-check is performed at the output of each stage. Record the current budget. Occupied:
[0225]
[0226] when The amplitude utilization factor is calculated according to the time. By scaling proportionally, the energy splitting parameters and all phases remain unchanged, thus ensuring that the budget constraints are strictly met.
[0227] In this embodiment, the K-stage processing network was trained before step 4.
[0228] The training objective function of the K-stage processing network is in the form of a hierarchical weighted sum:
[0229]
[0230]
[0231] in, For the number of training samples, This represents the total number of time slots for the drone. For the number of ground users, The normalization constant is for The stage processes all learnable parameters of the network. For the first In the training samples, the th Individual users in time slots Information freshness, For the first In the training samples, the th Individual users in time slots The package is available with instructions. For the first In the training samples, the th Individual users in time slots No. The smooth rate threshold indication for the stage.
[0232] Temperature coefficient for each stage is calculated as follows: Parameterization, where, For each stage of learnable scalar, initialize to 1.0 and prune to... ; According to the cosine curve attenuation rule:
[0233]
[0234] in, and These represent the upper and lower temperature limits, respectively, which can be set to 5 and 0.1. For the total number of training rounds, This is the current training round. (Regarding the rate threshold...) Adopting a step-by-step improvement rule from low to high, in the first... During each training round, the number of cycles is increased according to a linear rule to ensure stable convergence.
[0235] Step 5: Calculate the base station path rate and airborne receiver path rate based on the configuration parameters, port activation vector, and UAV's two-dimensional position output in Step 4. Determine whether the task is successful based on the calculated base station path rate and airborne receiver path rate, and update the information freshness. After the multi-timeslot process ends, output the system's average information freshness and corresponding transmission strategy.
[0236] The updated reflection configuration matrix based on the output of the last stage of the K-stage processing network in the current time slot. Transmission configuration matrix and port activation vector Two-dimensional position of the drone Calculate the base station path rate using the effective channel and non-orthogonal multiple access uplink receive model from step 2. and drone path rate According to the dual-path rate threshold rule Determine whether the user's task was successful, and update the information freshness status of all users according to the aforementioned information freshness recursion rule.
[0237] exist After each time slot rollover ends, the system calculates the average information freshness of all users' information across all time slots. And calculate the worst user information freshness. And user fairness metrics :
[0238]
[0239] in, For the first The average information freshness of each user in this inference. Output the system's average information freshness. The worst user information freshness, fairness indicators, and corresponding transmission strategies (panel configuration matrix for each time slot, port activation subset, and UAV two-dimensional position) are used as the final inference results.
[0240] To verify the technical effectiveness of the antenna-assisted low-latency air-to-space transmission method (the method of this invention) provided by this invention, the performance of the method of this invention and six comparative methods was simulated and verified in a UAV inspection scenario. The six comparative methods are: alternating optimization method (comparison method 1), pointer network port selection method (comparison method 2), random port selection method (comparison method 3), globally unified energy splitting method (comparison method 4), passive reconfigurable transmission reflector panel method (comparison method 5), and passive panel fixed port method (comparison method 6).
[0241] Among them, Comparison Method 1 is to alternate between solving the panel-track block and the port activation block: In each iteration, the port activation vector is first fixed, and the panel configuration and the UAV's two-dimensional position are jointly optimized; then the panel configuration and the UAV's two-dimensional position are fixed, and the port activation vector is optimized; the two sub-problems are solved by calling the convex optimization solver respectively, and the loop continues until the preset outer maximum iteration round.
[0242] Comparison Method 2 uses a pointer network for receiving port selection: the feature vectors of each candidate port are input into the encoder and decoder of the pointer network, and the decoder outputs the index of the selected port step by step until a preset number of ports are selected.
[0243] The difference between Comparison Method 3 and the method of the present invention is that Comparison Method 3 does not use graph attention and differentiable port selection modules, but randomly selects a preset number of ports from the candidate port set as an active port subset in each time slot. The panel configuration and the update of the UAV's two-dimensional position are still performed in the manner of the method of the present invention.
[0244] The difference between Comparison Method 4 and the method of the present invention is that Comparison Method 4 does not use element-by-element energy splitting parameters, but sets uniform energy splitting parameters for all elements of the panel, that is, the reflection and transmission power distribution ratio of all elements is the same; the amplitude utilization factor is still updated in an element-by-element manner.
[0245] The difference between Comparison Method 5 and the method of the present invention is that Comparison Method 5 uses a passive reconfigurable transmission reflection panel. Each component of the panel does not contain a power amplifier, and the reflection amplitude and transmission amplitude do not exceed 1, so there is no amplifier noise. The panel configuration parameters and the update of the two-dimensional position of the UAV are still performed in the manner of the method of the present invention.
[0246] The difference between Comparison Method 6 and the method of the present invention is that Comparison Method 6 uses a passive reconfigurable transmission reflection panel, does not configure the panel, fixes the reflection amplitude and transmission amplitude of each element of the panel to 1, fixes the phase to 0, fixes the port activation vector to a preset subset of ports selected at equal intervals from the candidate port set, and uses the trajectory obtained by the method of the present invention for the two-dimensional position of the UAV.
[0247] The simulation parameters are set as follows:
[0248] Ground users Take 5, the total number of time slots for the drone. Take 50 as the drone's flight altitude. Take 50 meters as an example, the maximum flight speed of the drone. Take 20 m / s. Number of panel components. Take 64, where and All values are set to 8; maximum hardware amplification per component. Take 5; Panel transmit power budget Take 0.5W. Number of candidate ports. Take 64, where and All values are 8; Number of activated ports Choose 6. The carrier frequency is 2.4 GHz, and the ground user transmit power... Take 30dBm as the base station receiver thermal noise power. Take -100dBm, airborne receiver thermal noise power Take -100 dBm, packet arrival probability Take 0.8, path loss exponent Take 2.8, the drone decoding capacity The total number of stages K in the K-stage processing network is set to 5. The K-stage processing network is trained offline using 50,000 training samples and 10,000 validation samples. The method of this invention uses the same simulation parameters as the comparative methods. The method of this invention and comparative methods 1 to 5 are evaluated on the same set of 100 independent test samples, each test sample containing 50 time slots.
[0249] Figure 2 The average information freshness of the system under different rate thresholds is demonstrated for the method of the present invention and comparative methods 1 to 5. Figure 2 It can be seen that as the rate threshold increases, the average information freshness of the system in the method of the present invention is always lower than that of the comparative methods, indicating that the method of the present invention optimizes the channel transmission path reconstruction and the receiver port reconstruction together, so that the gains of the two complement each other, and the performance is better than the scheme that relies on only one side design.
[0250] To evaluate the solution delay, the average solution time of the method of the present invention and the comparative method 1 on 10 test samples was statistically analyzed at a rate threshold of 16 bps / Hz. The results are shown in Table 1.
[0251] Table 1 Solution time of the method of the present invention and comparative method 1
[0252]
[0253] As can be seen from Table 1, the average solution time per sample of the method of the present invention is significantly less than that of the comparative method 1. Compared with the comparative method 1, the method of the present invention significantly reduces the solution delay.
[0254] Figures 3 to 6 The scheduling results of successful transmission by ground users via base station path and airborne receiver path in each time slot are shown for the same test sample with a rate threshold of 16bps / Hz, using the method of the present invention, comparative method 1, comparative method 3 and comparative method 6. Figures 3 to 6 In the table, a solid black block indicates that the ground user successfully transmitted via the base station path in the corresponding time slot; a block filled with diagonal lines indicates that the ground user successfully transmitted via the airborne receiver path in the corresponding time slot; and a blank space indicates that the ground user did not successfully transmit in the corresponding time slot. The simulation results of the method of this invention and various comparative methods in terms of user scheduling are shown in Table 2. The number of successful transmissions via the base station path, The number of successful transmissions via the airborne receiver path. This represents the system average information freshness of the test sample.
[0255] Table 2. User scheduling simulation results of the method of the present invention and the comparative method.
[0256]
[0257] from Figures 3 to 6 As shown in Table 2, the method of this invention provides the most balanced scheduling of base station paths, airborne receiver paths, and ground users, ensuring the most equal opportunity for each ground user to be scheduled. Furthermore, the method of this invention exhibits the lowest average system information freshness among the four methods. This indicates that the method of this invention, while reducing the average system information freshness, also maintains scheduling fairness among ground users.
[0258] The specific embodiments of the present invention enable those skilled in the art to understand or implement the invention. Various modifications to the above embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention.
[0259] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.
Claims
1. A method for antenna-assisted low-latency space-air transmission, characterized in that, Includes the following steps: Step 1: Obtain the initial system parameters and construct a data transmission system model based on the initial system parameters. The initial system parameters include at least the UAV's flight altitude, the UAV's flight area, the array size of the active reconfigurable transmission reflector, the transmit power budget of the active reconfigurable transmission reflector, the number of candidate ports of the reconfigurable receiving antenna array, the number of active ports of the reconfigurable receiving antenna array, the number of ground users, the rate threshold, and channel statistics parameters. Step 2: Construct the direct channel from the ground user to the base station, the reflection channel from the ground user to the active reconfigurable transmission reflector panel, the cascaded reflection channel from the active reconfigurable transmission reflector panel to the base station, the transmission channel from the active reconfigurable transmission reflector panel to the airborne receiver, and the port space correlation matrix of the reconfigurable receiving antenna array. Based on non-orthogonal multiple access uplink reception, serial interference cancellation and minimum mean square error merging are performed on the base station path to obtain the base station path rate. The base station path includes a direct path from the ground user's transmitted signal to the base station via a direct channel, and a reflection path from the ground user's transmitted signal to the base station via a reflection channel and a cascaded reflection channel. The signals from the direct path and the reflection path are superimposed at the base station. Serial interference cancellation is performed on the airborne receiver path to obtain the airborne receiver path rate. The path to the airborne receiver is the path from the ground user's transmitted signal to the airborne receiver via the reflection channel and the transmission channel; User information freshness is recursively deduced based on the base station path rate, airborne receiver path rate, data packet arrival indication, packet availability indication, and rate threshold, and the system average information freshness is calculated. Step 3: Using the configuration parameters of the active reconfigurable transmission reflector panel, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV as optimization variables, and minimizing the average information freshness of the system as the objective, a mixed integer nonconvex joint optimization problem is constructed under the constraints of parameter value range, UAV flight constraints, port selection constraints, and transmit power budget constraints of the active reconfigurable transmission reflector panel. Step 4: Based on the K-stage processing network, perform online block-based iterative solution in the outer and inner layers, and output the updated configuration parameters of the active reconfigurable transmission reflector panel, the port activation vector of the reconfigurable receiving antenna array, and the two-dimensional position of the UAV. The outer layer fixes the configuration parameters and port activation vector and jointly updates the two-dimensional position of the UAV in each time slot; the inner layer fixes the two-dimensional position of the UAV and updates the configuration parameters and port activation vector by time slot by the K-stage processing network. Step 5: For each time slot, calculate the base station path rate and airborne receiver path rate based on the updated configuration parameters, port activation vector, and UAV 2D position output in Step 4. Determine whether the task is successful based on the calculated base station path rate and airborne receiver path rate, and update the user information freshness. After all time slots are completed, output the system average information freshness and the corresponding transmission strategy.
2. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, In step 1, the channel statistical parameters include the thermal noise power of the base station receiver, the thermal noise power of the airborne receiver, the noise figure of the power amplifier per component, the Rice factor, and the path loss index; the system initial parameters also include the total number of UAV time slots, the energy splitting parameter, and the boundary parameters for the amplitude utilization factor; the data transmission system model is established in a unified coordinate system.
3. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, In step 3, the configuration parameters include reflection amplitude, transmission amplitude, energy splitting parameters, amplitude utilization factor, reflection phase, and transmission phase; The reflection amplitude and reflection phase constitute the reflection configuration matrix: The transmission amplitude and transmission phase constitute the transmission configuration matrix: In the formula, For active reconfigurable transmission reflective panels in time slots The reflection configuration matrix, For diagonal matrix operations, For the first Each component in the time slot The amplitude of reflection, , For the first Each component in the time slot The reflection phase, For active reconfigurable transmission reflective panels in time slots The transmission configuration matrix, For the first Each component in the time slot The transmission amplitude, , For the first Each component in the time slot The transmission phase, , For the array size of actively reconfigurable transmission reflective panels, The imaginary unit; For the first Each component in the time slot The energy splitting parameters, For the first Each component in the time slot Amplitude utilization factor, The upper limit of hardware amplification per component, These are the boundary parameters for the energy splitting parameter and the amplitude utilization factor.
4. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, In step 2, the signals from the direct path and the reflected path are superimposed at the base station and then received by the reconfigurable receiving antenna array. The reconfigurable receiving antenna array selects the effective channel of the base station path corresponding to the active port from the full-dimensional channel of the base station through a selection matrix. When the base station path performs serial interference cancellation and minimum mean square error merging, the interference plus noise covariance matrix is composed of the base station receiver thermal noise covariance and color noise covariance. The color noise covariance is formed by the power amplifier noise of each element. The power amplifier noise of each element is modulated by the reflection configuration matrix and then propagated to the base station through the cascaded reflection channel.
5. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, Step 4, the process of updating the UAV's two-dimensional position in the outer layer includes: Performing a first-order Taylor expansion on the path loss term yields a lower-bound affine function for the UAV's two-dimensional position: In the formula, In time slot No. The distance from a ground user to an active reconfigurable transmission reflector panel. In time slot The distance from the active reconfigurable transmission reflector panel to the base station, For the outer iteration round, For the first The first iteration of the outer layer The distance from a ground user to an active reconfigurable transmission reflector panel. For the first The distance from the active reconfigurable transmission reflector panel to the base station during the outer layer iteration process. For the first The slack variable of the distance from a ground user to the drone. Let the distance from the drone to the base station be a slack variable. The path loss index; Introduction and As a second-order cone constraint: In the formula, For the first The location of each ground user. For drones in time slots Two-dimensional position, The altitude at which the drone flies. Location of the base station; A weighted objective function is constructed using the square of the user information freshness as a weighting coefficient. The upper limit constraint of UAV displacement and the UAV flight area constraint are added to form a second-order cone programming subproblem and solve it to obtain the UAV's two-dimensional position after the outer layer iterative update.
6. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, It also includes training the K-stage processing network, constructing a supervision signal based on the task success flag, packet arrival indication, packet availability indication, and user information freshness, using the supervision signal as loss to train the learnable parameters of each stage, and freezing the network parameters after training is completed. The task success flag is generated based on the base station path rate, airborne receiver path rate, and rate threshold. The training objective function of the K-stage processing network adopts a hierarchical weighted sum form. For each stage, the smooth decision result of the output is weighted by the square of the freshness of user information and accumulated to obtain the loss of each stage. The total loss is calculated by summing the losses of each stage according to their stage numbers, applying increasing weights to each stage, and then normalizing the sum.
7. The antenna-assisted low-latency space-air transmission method according to claim 1, characterized in that, The K-stage processing network comprises K sequentially executed stages, with the output of each stage serving as the input to the next stage, and the output of the final stage serving as the inner-layer update result. The input to the first stage includes the initial port score, the UAV's two-dimensional position obtained from the outer-layer update, the configuration parameters obtained from the previous inner-layer update, the port activation vector obtained from the previous inner-layer update, and the channel state information of the current time slot. The update process of each stage sequentially includes a reflection phase and transmission phase alignment module, a stepwise energy splitting parameter and amplitude utilization factor projection gradient update module, a cross-block attention modulation module, and a graph attention and differentiable port selection module. Under the condition of fixing the reflection amplitude and transmission amplitude in the current stage input, the reflection phase and transmission phase alignment module constructs auxiliary variables for each ground user, constructs a service quality weighting with a learnable slope as a parameter, multiplies the service quality weighting with the auxiliary variables to obtain the phase alignment coefficient, and updates the reflection phase and transmission phase of each element according to the phase alignment coefficient. The stepwise energy splitting parameter and amplitude utilization factor projection gradient update module calculates the partial derivatives of the loss function with respect to the energy splitting parameter and amplitude utilization factor under the condition of fixed reflection phase and transmission phase. It calculates the pre-projection update values of the energy splitting parameter and amplitude utilization factor according to the learnable step size and the partial derivatives. The pre-projection update values of the energy splitting parameter and amplitude utilization factor are projected to their respective value ranges by the box projection operator to obtain the updated energy splitting parameter and amplitude utilization factor. The cross-block attention modulation module constructs panel state feature vectors, trajectory feature vectors, and port feature vectors. The three types of feature vectors are mapped to a shared latent space through independent linear projection and layer normalization, respectively, resulting in three tokens. The three tokens form a token matrix. Multi-head attention operation is applied to the token matrix. The output of the multi-head attention operation is linearly transformed, hyperbolic tangent activated, and multiplied by a preset upper limit to obtain the modulation bias and port score bias of the active reconfigurable transmission reflection panel. The modulation bias is fed back to the element-wise energy splitting parameter and amplitude utilization factor projection gradient update module. The port score bias is superimposed on the graph attention port score in the next stage. The graph attention and differentiable port selection module is used to construct an undirected graph containing port nodes, user nodes and global context nodes, assign an initial embedding to each node, and iteratively update the embedding of each node in a two-layer graph attention network. The embedding of the port node is selected from the updated node embeddings, and the score of each port is obtained through a multilayer perceptron. After superimposing the port score bias from the previous stage, a Gumbel perturbation is introduced to calculate soft selection, making the sum of the soft selection values equal to... Select the highest-rated port from all ports according to the hard selection rule. The set of active ports consists of several ports. Number of activated ports; By employing direct gradient estimation, the gradient corresponding to hard selection is backpropagated through the soft selection path.
