An Adaptive Optimization Method for UAV Wireless Links Based on Perception Fusion

Abstract

This invention discloses an adaptive optimization method for UAV wireless links based on perception fusion, belonging to the field of link optimization technology. The method includes: collecting multimodal environmental perception data from the UAV and performing time synchronization and spatial registration to construct a BEV semantic electromagnetic grid map; predicting future channel vectors based on the BEV semantic electromagnetic grid map using a temporal attention unfolding network; constructing a channel prediction correction model, a motion prediction model, and a link performance mapping model, which are then combined to form a digital twin state transition function; constructing state vectors, generating recommended actions through a PPO policy network, and generating candidate action sequences based on neighborhood sampling of the recommended actions; performing future time step pre-playing on the candidate action sequences using the digital twin state transition function, calculating the cumulative discount reward for each candidate action sequence based on the pre-playing results, and outputting the optimal action decision.

Description

Technical Field

[0001] This invention relates to the field of link optimization technology, and in particular to an adaptive optimization method for UAV wireless links based on perception fusion. Background Technology

[0002] With the research and standardization of sixth-generation mobile communication systems, UAV communication is regarded as a key component of the future integrated air-space-ground network due to its high mobility, flexible deployment and wide coverage capabilities. During UAV flight, the wireless link environment exhibits highly dynamic and non-stationary characteristics, with drastic changes in multipath fading, blockage and Doppler effects, leading to severe performance degradation of traditional link adaptation schemes based on ideal channel assumptions (such as fixed modulation and coding schemes and power control). In recent years, deep learning and reinforcement learning have been introduced into the field of wireless resource management, providing a new technical path for wireless link adaptation.

[0003] However, existing methods have the following shortcomings: First, the conversion process from environmental perception data to communication prior information is relatively crude. Most solutions directly perform 3D semantic segmentation on LiDAR point clouds or RGB images, resulting in high computational load and difficulty in meeting the 6G millisecond latency requirement on airborne edge computing platforms. Second, most channel prediction models are based solely on historical pilot observations, ignoring the strong causal constraints of environmental geometry and semantic categories on future channel evolution. Even when some work incorporates electromagnetic maps, it adopts simple feature stitching or adds regularization terms, failing to achieve deep coupling between perception priors and prediction models at the network structure design level. Third, link adaptive decision-making mechanisms are mostly single-step strategies (such as instantaneous optimal power allocation), lacking the ability to plan for the cumulative effects of multiple future steps, leading to lag in response to scenarios of rapid switching due to occlusion or sudden interference. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a UAV wireless link adaptive optimization method based on perception fusion, which solves the problems of low efficiency of environmental perception conversion, insufficient coupling between perception prior and channel prediction, and lack of multi-step planning in existing UAV link adaptation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an adaptive optimization method for UAV wireless links based on perception fusion, comprising:

[0008] Collect multimodal environmental perception data from UAVs and perform time synchronization and spatial registration to construct a semantic electromagnetic grid map of BEVs;

[0009] Based on the BEV semantic electromagnetic grid map, a temporal attention unfolding network is used to predict future channel vectors;

[0010] Construct a channel prediction correction model, a motion prediction model, and a link performance mapping model, and combine them to form a digital twin state transition function;

[0011] Construct a state vector, generate recommended actions through a PPO policy network, and generate candidate action sequences based on neighborhood sampling of the recommended actions.

[0012] The candidate action sequence is simulated for future time steps using a digital twin state transition function. Based on the simulation results, the cumulative discounted reward for each candidate action sequence is calculated, and the optimal action decision is output.

[0013] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the UAV multimodal environment perception data includes RGB images of the foreground environment, three-dimensional point cloud data, UAV GPS three-dimensional position, three-dimensional velocity, attitude angle, and corresponding timestamps.

[0014] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps for time synchronization and spatial registration are as follows:

[0015] Based on GPS timestamps, time synchronization is performed on RGB images and 3D point cloud data using linear interpolation and cubic spline interpolation, respectively.

[0016] After time synchronization is completed, a rotation matrix is ​​constructed from the sensor local coordinate system to the UAV body coordinate system using GPS attitude angles. This transforms the 3D point cloud data from the sensor coordinate system to the body coordinate system, resulting in a time-synchronized and spatially registered RGB image. and 3D point cloud data .

[0017] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps for constructing the BEV semantic electromagnetic grid map are as follows:

[0018] The time-synchronized and spatially registered RGB images and 3D point cloud data are converted into BEV raster representation in the UAV body coordinate system, and the current position of the UAV is defined as the center of the 2D raster region.

[0019] right Each point in ,calculate The grid index falls on each grid cell, and all point cloud information falling within it is aggregated to calculate the maximum height, point density, and average reflection intensity.

[0020] right A lightweight semantic segmentation network is used to output the semantic category of each pixel, which includes buildings, trees, ground, moving objects, and open airspace.

[0021] Pixels are back-mapped to the BEV grid using a camera projection model;

[0022] For each BEV raster cell, the semantic histogram of all pixels inside is calculated, and the mode is taken as the semantic category of the raster.

[0023] Each semantic category is assigned a set of frequency-related electromagnetic parameter vectors;

[0024] Divide the BEV grid along the vertical direction. Layers are used to obtain a 3D raster map. Each grid cell stores an obstacle marker, a penetration loss coefficient, and a scattering coefficient. The obstacle marker is determined by comparing the maximum height of the grid cell with the current flight altitude of the UAV, and the penetration loss coefficient is calculated from the electromagnetic parameter vector.

[0025] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the step of predicting future channel vectors based on BEV semantic electromagnetic grid maps and using a temporal attention unfolding network is as follows:

[0026] Acquire historical channel observation sequences and 3D raster map sequences;

[0027] The input feature matrix is ​​constructed based on historical channel observation sequences and 3D raster map sequences;

[0028] Construct a temporal attention unfolding network, which consists of an encoder, a Transformer Encoder, and a decoder;

[0029] The input feature matrix is ​​fed into the encoder. Each encoder layer performs temporal convolution, channel splitting convolution and feature concatenation step by step. The last encoder layer outputs the final compressed features.

[0030] The Transformer Encoder is used to perform self-attention processing on the final compressed features to obtain self-attention features;

[0031] The decoder is symmetric to the encoder. It upsamples through transposed convolution and introduces skip connections to concatenate the features of the corresponding layer of the encoder to the input of the decoder. The last decoder layer outputs the channel prediction vector for future time steps.

[0032] Define data fitting loss, physical consistency loss, Doppler consistency constraint, and spatial consistency loss, and jointly construct the total loss function;

[0033] The network is trained using a training set generated by a ray tracing simulator. The model is constrained by a total loss function, and gradient descent optimization is performed using the Adam optimizer to update the network parameters. If the network loss no longer decreases significantly during continuous iteration, the iteration is stopped and the network parameters are output.

[0034] The trained temporal attention unfolding network is used to generate channel prediction vectors for future time moments.

[0035] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps for constructing a channel prediction correction model, a motion prediction model, and a link performance mapping model, and then combining them to form a digital twin state transition function, are as follows.

[0036] Define the channel deviation, use a discrete Kalman filter to estimate the deviation state online, recursively update the Kalman filter to generate the filtered deviation estimate, use the deviation estimate to correct the channel prediction vector, and obtain the corrected complete channel prediction vector;

[0037] Predict future flight paths based on the current movement status of the drone;

[0038] Estimate SINR based on the corrected complete channel prediction vector and candidate actions;

[0039] Calculate the system's achievable throughput based on SINR;

[0040] The channel prediction correction model, motion prediction model, and link performance mapping model are integrated into a complete digital twin state transition function. The digital twin state transition function receives the current state and candidate actions, and outputs the state and performance indicators for the next time step.

[0041] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps of constructing the state vector and generating recommended actions through the PPO policy network are as follows:

[0042] Construct the complete state vector at the current moment;

[0043] Modulation coding scheme index, transmit power, beam index, and frequency band selection are used to construct action decision variables;

[0044] The PPO policy network, which uses a multilayer perceptron structure, takes the complete state vector at the current moment as input. For discrete actions, it outputs the probability distribution of the action space; for continuous actions, it outputs the mean and standard deviation of the Gaussian distribution. The discrete action with the highest probability and the mean of the Gaussian distribution of the continuous action are selected as the recommended actions.

[0045] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the step of generating candidate action sequences based on recommended action neighborhood sampling includes the following steps:

[0046] For discrete actions in the recommended actions, multinomial distribution sampling is performed with the discrete value corresponding to the discrete action as the center. For continuous actions in the recommended actions, Gaussian noise is added with the continuous action as the mean to obtain a set of candidate action sequences.

[0047] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps of performing future time-step pre-simulation on candidate action sequences using a digital twin state transition function and calculating the cumulative discount reward for each candidate action sequence based on the pre-simulation results are as follows:

[0048] For each candidate action sequence, a digital twin state transition function is used to simulate the state transition in the next H steps, and the process is iterated step by step.

[0049] For each candidate action sequence, calculate the cumulative discount reward based on the iteration results.

[0050] As a preferred embodiment of the UAV wireless link adaptive optimization method based on perception fusion described in this invention, the steps for outputting the optimal action decision are as follows:

[0051] The first action in the candidate action sequence with the largest cumulative reward is selected as the optimal action decision. The optimal action decision is then decoded into specific physical layer configuration parameters and distributed.

[0052] The beneficial effects of this invention are as follows: By transforming high-dimensional, heterogeneous multimodal sensing data (images, point clouds, poses) into compact, physically semantic 3D raster representations, low-latency, high-fidelity representation of perceptual priors is achieved; by using a temporal attention unfolding network to achieve deep coupling between perceptual priors and channel prediction, the limitations of existing methods that rely solely on historical channels or simple electromagnetic graph splicing are overcome; by constructing a digital twin state transition function that includes channel correction, motion prediction, and performance mapping, online modeling and self-correction of dynamic environmental evolution are achieved; by generating recommended actions through PPO and constructing candidate sequences by combining neighborhood sampling, an effective combination of reinforcement learning strategies and multi-step search is achieved, striking a balance between search efficiency and optimality; by selecting the action sequence with the maximum cumulative reward over multiple steps through digital twin pre-playing, link adaptation is upgraded from reactive decision-making to forward-looking planning, making it more adaptable to complex dynamic environments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of an adaptive optimization method for UAV wireless links based on perception fusion, as shown in Example 1.

[0055] Figure 2 This is a flowchart of the BEV semantic electromagnetic grid map construction process in Example 1.

[0056] Figure 3 The flowchart for constructing the digital twin state transition function in Example 1 is shown.

[0057] Figure 4 This is a flowchart of the decision optimization process in Example 1. Detailed Implementation

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0060] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0061] Example 1, referring to Figure 1 , Figure 2 , Figure 3 and Figure 4 This is the first embodiment of the present invention, which provides an adaptive optimization method for UAV wireless links based on perception fusion, including the following steps:

[0062] S1. Collect multimodal environmental perception data of UAVs and perform time synchronization and spatial registration to construct a BEV semantic electromagnetic grid map;

[0063] Specifically, the UAV multimodal environmental perception data includes RGB images of the foreground environment, 3D point cloud data, UAV GPS 3D position, 3D velocity, attitude angle, and corresponding timestamps;

[0064] Furthermore, based on GPS timestamps, time synchronization is performed on RGB images and 3D point cloud data using linear interpolation and cubic spline interpolation, respectively.

[0065] After time synchronization is completed, a rotation matrix is ​​constructed from the sensor local coordinate system to the UAV body coordinate system using GPS attitude angles. This transforms the 3D point cloud data from the sensor coordinate system to the body coordinate system, resulting in a time-synchronized and spatially registered RGB image. and 3D point cloud data ;

[0066] Furthermore, the time-synchronized and spatially registered RGB images and 3D point cloud data are converted into BEV raster representation in the UAV body coordinate system, and the current position of the UAV is defined as the center of the 2D raster region;

[0067] right Each point in ,calculate The grid index falls on each grid cell, and all point cloud information falling within it is aggregated to calculate the maximum height, point density, and average reflection intensity.

[0068] right A lightweight semantic segmentation network is used to output the semantic category of each pixel, which includes buildings, trees, ground, moving objects, and open airspace.

[0069] The pixels are inversely mapped to the BEV raster using the camera projection model, expressed as:

[0070] ;

[0071] ;

[0072] in, Represents pixels The three-dimensional spatial direction vector, and Represents pixels Image coordinates, Represents the camera intrinsic parameter matrix. Represents pixels The coordinates of the projected position on the BEV grid. Indicates time The horizontal component of the three-dimensional position of the UAV Represents the scaling factor numerator. This indicates the drone's current altitude above the ground;

[0073] For each BEV raster cell, the semantic histogram of all pixels inside is calculated, and the mode is taken as the semantic category of the raster.

[0074] Each semantic category is assigned a set of frequency-related electromagnetic parameter vectors, expressed as follows:

[0075] ;

[0076] in, Represents an electromagnetic parameter vector. Represents the relative permittivity. Indicates electrical conductivity. This represents the surface roughness factor, with a value range of [value range missing]. The initial parameters were obtained through literature review and laboratory calibration of typical materials;

[0077] Divide the BEV grid along the vertical direction. Layers are used to obtain a 3D raster map. Each grid cell stores an obstacle marker, a penetration loss coefficient, and a scattering coefficient. The obstacle marker is determined by comparing the maximum height of the grid cell with the current flight altitude of the UAV, and the penetration loss coefficient is calculated from the electromagnetic parameter vector.

[0078] The penetration loss coefficient is expressed as:

[0079] ;

[0080] in, Indicates the penetration loss coefficient. Indicates the operating frequency. Represents an electromagnetic parameter vector. Indicates grid thickness. This represents the fitting constant for the dielectric constant.

[0081] The expression for the scattering coefficient is:

[0082] ;

[0083] in, Indicates the scattering coefficient;

[0084] It should be noted that BEV refers to bird's-eye view, and BEV grid is a spatial data representation structure that projects three-dimensional physical space onto a two-dimensional horizontal ground and discretizes it at equal intervals.

[0085] S2. Based on the BEV semantic electromagnetic grid map, predict future channel vectors using a temporal attention unfolding network.

[0086] Specifically, historical channel observation sequences and three-dimensional raster map sequences are acquired;

[0087] The input feature matrix is ​​constructed based on historical channel observation sequences and 3D raster map sequences, and its expression is:

[0088] ;

[0089] in, Represents the input feature matrix. Represents the historical channel observation sequence. This indicates that the 3D raster map exhibits are represented as one-dimensional vectors;

[0090] Construct a temporal attention unfolding network, which consists of an encoder, a Transformer Encoder, and a decoder;

[0091] The input feature matrix is ​​fed into the encoder. Each encoder layer performs temporal convolution, channel splitting convolution and feature concatenation step by step. The last encoder layer outputs the final compressed features.

[0092] The Transformer Encoder is used to perform self-attention processing on the final compressed features to obtain self-attention features;

[0093] The decoder is symmetric to the encoder. It upsamples through transposed convolution and introduces skip connections to concatenate the features of the corresponding layer of the encoder to the input of the decoder. The last decoder layer outputs the channel prediction vector for future time steps.

[0094] Define data fitting loss, physical consistency loss, Doppler consistency constraint, and spatial consistency loss, and jointly construct the total loss function;

[0095] The network is trained using a training set generated by a ray tracing simulator. The model is constrained by a total loss function, and gradient descent optimization is performed using the Adam optimizer to update the network parameters. If the network loss no longer decreases significantly during continuous iteration, the iteration is stopped and the network parameters are output.

[0096] The trained temporal attention unfolding network is used to generate channel prediction vectors for future time moments.

[0097] S3. Construct a channel prediction correction model, a motion prediction model, and a link performance mapping model, and combine them to form a digital twin state transition function;

[0098] Specifically, the channel deviation is defined, the state of the deviation is estimated online using a discrete Kalman filter, the filtered deviation estimate is generated by recursively updating the Kalman filter, and the channel prediction vector is corrected using the deviation estimate to obtain the corrected complete channel prediction vector.

[0099] Based on the current motion state of the drone, the future trajectory can be predicted using the following expression:

[0100] ;

[0101] in, Indicating future moments of drones The three-dimensional position vector, Indicates the prediction time offset. Indicates the drone at a certain time The three-dimensional velocity vector, Indicates the acceleration of the drone. The term represents the random wind disturbance, which follows a zero-mean Gaussian distribution. The covariance matrix is ​​calculated from the real-time measurements of the airborne anemometer.

[0102] Based on the corrected complete channel prediction vector and candidate actions, the SINR is estimated using the following expression:

[0103] ;

[0104] in, Indicating drones in the future The predicted signal-to-interference-plus-noise ratio, Represents the revised future moment The complete channel prediction vector, This represents the conjugate transpose operation. Indicates future time Beamforming vectors corresponding to candidate actions, based on future The candidate actions are selected from the preset DFT codebook. This represents the beamforming vector corresponding to other simultaneously transmitted co-channel interference streams. The noise power at the receiving end is determined by the receiver noise figure and the system bandwidth.

[0105] The achievable throughput is predicted based on SINR using a modified Shannon formula.

[0106] The channel prediction correction model, motion prediction model, and link performance mapping model are integrated into a complete digital twin state transition function. The digital twin state transition function receives the current state and candidate actions, and outputs the state and performance indicators for the next time step.

[0107] S4. Construct a state vector, generate recommended actions through the PPO policy network, and generate candidate action sequences based on neighborhood sampling of recommended actions.

[0108] Specifically, the complete state vector at the current moment is constructed as follows:

[0109] ;

[0110] in, Represents the complete state vector. This represents the corrected complete channel prediction vector. Indicates the amount of data to be sent;

[0111] Modulation coding scheme index, transmit power, beam index, and frequency band selection are used to construct action decision variables;

[0112] The PPO policy network uses a multilayer perceptron structure. It takes the complete state vector at the current moment as input, outputs the probability distribution of the action space for discrete actions, and outputs the mean and standard deviation of the Gaussian distribution for continuous actions. It selects the discrete action with the highest probability and the mean of the Gaussian distribution of the continuous action as the recommended action.

[0113] The PPO policy network was trained in a simulation environment during the offline phase;

[0114] Furthermore, for discrete actions in the recommended actions, multinomial distribution sampling is performed with the discrete value corresponding to the discrete action as the center; for continuous actions in the recommended actions, Gaussian noise is added with the continuous action as the mean to obtain a set of candidate action sequences.

[0115] It should be noted that PPO refers to the Proximal Policy Optimization Reinforcement Learning Algorithm.

[0116] S5. Use the digital twin state transition function to perform a future time step preview of the candidate action sequence, calculate the cumulative discount reward for each candidate action sequence based on the preview results, and output the optimal action decision;

[0117] Specifically, for each candidate action sequence, a digital twin state transition function is used to simulate the state transition in the next H steps, and the process is iterated step by step.

[0118] For each candidate action sequence, calculate the cumulative discount reward based on the iteration results, expressed as:

[0119] ;

[0120] ;

[0121] in, This represents the cumulative discount reward for the k-th candidate action sequence. This represents the discount factor; in this invention, the exemplary value is 0.95. This represents the single-step reward function. , and The weighting coefficients are determined through cross-validation. This indicates that the k-th candidate action sequence is at time [time]. The achievable throughput predicted by the digital twin state transition function is as follows. This represents the system's theoretical maximum throughput. This indicates that the k-th candidate action sequence is at time [time]. The set transmit power value, Indicates the maximum allowable transmit power of the system. Indicates time The interrupt indicator flag is a binary value with a range of values. If the digital twin state transition function at time 1 If the predicted SINR is lower than the preset interruption threshold, then The value is 1 if it is not 0 otherwise. The preset interruption threshold is determined by the target bit error rate and the adjustment threshold corresponding to the current modulation and coding scheme.

[0122] Furthermore, the first action in the candidate action sequence with the largest cumulative reward is selected as the optimal action decision, and the optimal action decision is decoded into specific physical layer configuration parameters and issued.

[0123] This embodiment also provides a computer device applicable to a perceptual fusion-based adaptive optimization method for UAV wireless links, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the perceptual fusion-based adaptive optimization method for UAV wireless links as proposed in the above embodiment.

[0124] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0125] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements an adaptive optimization method for UAV wireless links based on perception fusion as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0126] In summary, this invention achieves low-latency, high-fidelity representation of perceptual priors by transforming high-dimensional, heterogeneous multimodal sensing data (images, point clouds, poses) into compact, physically semantic 3D raster representations. It overcomes the limitations of existing methods that rely solely on historical channels or simple electromagnetic mapping by using a temporal attention unfolding network to achieve deep coupling between perceptual priors and channel prediction. Furthermore, it enables online modeling and self-correction of dynamic environmental evolution by constructing a digital twin state transition function that includes channel correction, motion prediction, and performance mapping. Finally, it effectively combines reinforcement learning strategies with multi-step search by generating recommended actions through PPO and constructing candidate sequences using neighborhood sampling, achieving a balance between search efficiency and optimality. Finally, it selects the action sequence with the maximum cumulative reward across multiple steps through digital twin pre-playing, elevating link adaptation from reactive decision-making to forward-looking planning, making it more adaptable to complex dynamic environments.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive optimization method for UAV wireless links based on perception fusion, characterized in that: include, Collect multimodal environmental perception data from UAVs and perform time synchronization and spatial registration to construct a semantic electromagnetic grid map of BEVs; Based on the BEV semantic electromagnetic grid map, a temporal attention unfolding network is used to predict future channel vectors; Construct a channel prediction correction model, a motion prediction model, and a link performance mapping model, and combine them to form a digital twin state transition function; Construct a state vector, generate recommended actions through a PPO policy network, and generate candidate action sequences based on neighborhood sampling of the recommended actions. The candidate action sequence is simulated for future time steps using a digital twin state transition function. Based on the simulation results, the cumulative discounted reward for each candidate action sequence is calculated, and the optimal action decision is output.

2. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 1, characterized in that: The UAV's multimodal environmental perception data includes RGB images of the foreground environment, 3D point cloud data, UAV GPS 3D position, 3D velocity, attitude angle, and corresponding timestamps.

3. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 2, characterized in that: The steps for time synchronization and spatial registration are as follows. Based on GPS timestamps, time synchronization is performed on RGB images and 3D point cloud data using linear interpolation and cubic spline interpolation, respectively. After time synchronization is completed, a rotation matrix is ​​constructed from the sensor local coordinate system to the UAV body coordinate system using GPS attitude angles. This transforms the 3D point cloud data from the sensor coordinate system to the body coordinate system, resulting in a time-synchronized and spatially registered RGB image. and 3D point cloud data .

4. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 3, characterized in that: The steps for constructing the BEV semantic electromagnetic grid map are as follows: The time-synchronized and spatially registered RGB images and 3D point cloud data are converted into BEV raster representation in the UAV body coordinate system, and the current position of the UAV is defined as the center of the 2D raster region. right Each point in ,calculate The grid index falls on each grid cell, and all point cloud information falling within it is aggregated to calculate the maximum height, point density, and average reflection intensity. right A lightweight semantic segmentation network is used to output the semantic category of each pixel, which includes buildings, trees, ground, moving objects, and open airspace. Pixels are back-mapped to the BEV grid using a camera projection model; For each BEV raster cell, the semantic histogram of all pixels inside is calculated, and the mode is taken as the semantic category of the raster. Each semantic category is assigned a set of frequency-related electromagnetic parameter vectors; Divide the BEV grid along the vertical direction. Layers are used to obtain a 3D raster map. Each grid cell stores an obstacle marker, a penetration loss coefficient, and a scattering coefficient. The obstacle marker is determined by comparing the maximum height of the grid cell with the current flight altitude of the UAV, and the penetration loss coefficient is calculated from the electromagnetic parameter vector.

5. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 4, characterized in that: The method for predicting future channel vectors based on BEV semantic electromagnetic grid maps using a temporal attention unfolding network is as follows: Acquire historical channel observation sequences and 3D raster map sequences; The input feature matrix is ​​constructed based on historical channel observation sequences and 3D raster map sequences; Construct a temporal attention unfolding network, which consists of an encoder, a Transformer Encoder, and a decoder; The input feature matrix is ​​fed into the encoder. Each encoder layer performs temporal convolution, channel splitting convolution and feature concatenation step by step. The last encoder layer outputs the final compressed features. The Transformer Encoder is used to perform self-attention processing on the final compressed features to obtain self-attention features; The decoder is symmetric to the encoder. It upsamples through transposed convolution and introduces skip connections to concatenate the features of the corresponding layer of the encoder to the input of the decoder. The last decoder layer outputs the channel prediction vector for future time steps. Define data fitting loss, physical consistency loss, Doppler consistency constraint, and spatial consistency loss, and jointly construct the total loss function; The network is trained using a training set generated by a ray tracing simulator. The model is constrained by a total loss function, and gradient descent optimization is performed using the Adam optimizer to update the network parameters. If the network loss no longer decreases significantly during continuous iteration, the iteration is stopped and the network parameters are output. The trained temporal attention unfolding network is used to generate channel prediction vectors for future time moments.

6. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 5, characterized in that: The steps for constructing the channel prediction correction model, motion prediction model, and link performance mapping model, and combining them to form a digital twin state transition function, are as follows. Define the channel deviation, use a discrete Kalman filter to estimate the deviation state online, recursively update the Kalman filter to generate the filtered deviation estimate, use the deviation estimate to correct the channel prediction vector, and obtain the corrected complete channel prediction vector; Predict future flight paths based on the current movement status of the drone; Estimate SINR based on the corrected complete channel prediction vector and candidate actions; Calculate the system's achievable throughput based on SINR; The channel prediction correction model, motion prediction model, and link performance mapping model are integrated into a complete digital twin state transition function. The digital twin state transition function receives the current state and candidate actions, and outputs the state and performance indicators for the next time step.

7. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 6, characterized in that: The process of constructing the state vector and generating recommended actions through the PPO policy network is as follows: Construct the complete state vector at the current moment; Modulation coding scheme index, transmit power, beam index, and frequency band selection are used to construct action decision variables; The PPO policy network, which uses a multilayer perceptron structure, takes the complete state vector at the current moment as input. For discrete actions, it outputs the probability distribution of the action space; for continuous actions, it outputs the mean and standard deviation of the Gaussian distribution. The discrete action with the highest probability and the mean of the Gaussian distribution of the continuous action are selected as the recommended actions.

8. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 7, characterized in that: The steps for generating candidate action sequences based on neighborhood sampling of recommended actions are as follows: For discrete actions in the recommended actions, multinomial distribution sampling is performed with the discrete value corresponding to the discrete action as the center. For continuous actions in the recommended actions, Gaussian noise is added with the continuous action as the mean to obtain a set of candidate action sequences.

9. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 8, characterized in that: The steps for performing future time-step simulations on candidate action sequences using digital twin state transition functions and calculating the cumulative discounted reward for each candidate action sequence based on the simulation results are as follows. For each candidate action sequence, a digital twin state transition function is used to simulate the state transition in the next H steps, and the process is iterated step by step. For each candidate action sequence, calculate the cumulative discount reward based on the iteration results.

10. The UAV wireless link adaptive optimization method based on perception fusion as described in claim 9, characterized in that: The steps for determining the optimal action decision are as follows: The first action in the candidate action sequence with the largest cumulative reward is selected as the optimal action decision. The optimal action decision is then decoded into specific physical layer configuration parameters and distributed.

Images