Unmanned aerial vehicle communication system based on machine learning and intelligent reflector assistance

Abstract

This invention discloses a UAV communication system, method, and storage medium based on machine learning and intelligent reflector assistance. The system includes a UAV platform, a ground base station, an intelligent reflector, and user equipment; it also includes a multimodal perception fusion module for collecting and fusing multi-source heterogeneous environmental information, extracting spatiotemporal features through attention mechanisms or graph neural networks, and generating a high-dimensional state representation vector; and a machine learning decision module for receiving this vector and generating a joint control strategy end-to-end based on a deep neural network model, including UAV flight attitude adjustment commands and intelligent reflector phase displacement matrix control commands. This invention achieves joint optimization of the UAV trajectory and intelligent reflector phase through multimodal perception fusion and end-to-end machine learning decision-making, significantly improving the performance, environmental adaptability, and energy efficiency of the communication system.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a UAV communication system, method, and storage medium based on machine learning and intelligent reflective surface assistance. Background Technology

[0002] Unmanned Aerial Vehicles (UAVs) have demonstrated great potential in emergency communications, temporary hotspot coverage, and IoT data collection due to their high mobility, rapid deployment, and on-demand line-of-sight links. However, in complex urban environments, signal attenuation caused by building obstruction and frequent link interruptions remain key factors limiting their performance.

[0003] Intelligent Reflecting Surfaces (IRS), as an emerging passive repeater technology, consist of a large number of low-cost, programmable reflective elements. By intelligently controlling the phase and / or amplitude of the incident signal, they can enhance the signal and suppress interference, thereby significantly improving communication quality. Introducing intelligent reflective surfaces into UAV communication systems is expected to combine the advantages of both technologies.

[0004] In the existing technology, although there are schemes that use reinforcement learning to optimize the trajectory of UAVs (such as CN110488861B) and schemes that optimize the phase of intelligent reflectors based on channel statistics (such as CN113644940A), these schemes are limited to single-point optimization or step-by-step optimization.

[0005] First, existing solutions lack a deep fusion mechanism for multi-source heterogeneous sensing information. They often simply splice together UAV inertial data and channel information, making it difficult to uncover the deep coupling characteristics of the environment in the spatiotemporal dimension, resulting in insufficient state representation capabilities.

[0006] Secondly, existing solutions typically decompose UAV trajectory planning and reflector phase shift design into alternating optimization sub-problems, ignoring the strong coupling and mutual influence between the two in dynamic environments, making it difficult to achieve end-to-end global optimal mapping from initial perception to final control commands.

[0007] Furthermore, the existing system architecture cannot adaptively switch between fixed intelligent reflector assistance and airborne intelligent reflector modes based on real-time link congestion levels, drone remaining energy, and service priorities. This results in a significant decrease in system robustness and energy efficiency when faced with sudden occlusion or energy consumption constraints.

[0008] Existing UAV communication optimization methods are mostly single-variable optimizations, lacking joint modeling of channel models, trajectory planning, phase modulation, and power allocation; machine learning applications are mostly limited to general descriptions, without constructing calculable and verifiable system models and quantitative evaluation systems, resulting in poor robustness, insufficient real-time performance, and difficulty in ensuring communication reliability and energy efficiency in dynamic environments.

[0009] Therefore, there is an urgent need for a system that can organically combine multimodal perception fusion, end-to-end joint decision-making, and adaptive mode switching to achieve adaptive communication in dynamic environments. Summary of the Invention

[0010] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a UAV communication system, method and storage medium based on machine learning and intelligent reflective surface assistance. By introducing multimodal spatiotemporal feature fusion, end-to-end closed-loop decision-making, online incremental learning and adaptive cooperative mode switching mechanism, the invention solves the problems of low decision quality, poor environmental adaptability and insufficient energy efficiency in the prior art.

[0011] To achieve the above objectives, the first aspect of the present invention provides a UAV communication system based on machine learning and intelligent reflective surface assistance, employing the following technical solution:

[0012] A UAV communication system based on machine learning and intelligent reflector assistance includes:

[0013] Unmanned aerial vehicle platform, ground base station, smart reflector and at least one user device;

[0014] The unmanned aerial vehicle platform is equipped with a communication module and a flight control module;

[0015] The intelligent reflective surface is equipped with multiple controllable reflective units for phase and / or amplitude modulation of incident wireless signals.

[0016] The system also includes a multimodal perception fusion module, which is used to collect and fuse multi-source heterogeneous environmental information, including the state data of the UAV platform, channel state information, user equipment demand information, and environmental interference information. The system also extracts spatiotemporal features from the fused information through an attention mechanism or graph neural network to generate a high-dimensional state representation vector.

[0017] The machine learning decision module is used to receive the high-dimensional state representation vector and generate a joint control strategy end-to-end based on a deep neural network model. The joint control strategy includes at least flight attitude adjustment commands for the UAV platform and control commands for the reflection phase shift matrix of the intelligent reflector.

[0018] The adaptive switching module for collaboration mode is used to dynamically select or mix at least two preset UAV-intelligent reflector collaboration modes based on real-time environmental information. The collaboration modes include a first collaboration mode and a second collaboration mode, wherein the first collaboration mode is a fixed deployment auxiliary mode for intelligent reflectors and the second collaboration mode is an airborne mode for intelligent reflectors.

[0019] The invention also includes a channel modeling unit for constructing a cascaded channel representation model of base station-intelligent reflector-UAV-user, providing physical layer model support for the system; the machine learning decision module implements policy learning based on the Markov decision process model and generates a theoretical representation of the system state of link performance, energy efficiency and decision adaptability, so that perception, modeling, decision and reconstruction form a complete closed loop.

[0020] It should be noted that the core processing unit of the present invention (including the multimodal perception fusion module, machine learning decision module and collaborative mode adaptive switching module) can be deployed in the edge computing unit carried by the cloud server, the ground base station, or the UAV platform. These different physical deployment methods are all equivalent transformations of the present invention and should all fall within the protection scope of the present invention.

[0021] Furthermore, in a preferred embodiment of the present invention, the machine learning decision module includes:

[0022] The strategy generation submodule uses a deep reinforcement learning network or a multi-task deep learning network to output the flight attitude adjustment command and the reflection phase shift matrix control command in real time according to the preset reward function.

[0023] The reward function includes at least a weighted factor for multiple objectives, such as communication throughput, negative link interruption probability, drone energy consumption, and service fairness.

[0024] Preferably, the reward function is specifically expressed as:

[0025] ;

[0026] in:

[0027] express Instant reward value at any given moment;

[0028] express All moments The sum of the instantaneous data transmission rates of each user device;

[0029] express The estimated probability of system link interruption at any given time, which can be estimated based on real-time channel quality and a preset interruption threshold;

[0030] express Estimated energy consumption of the drone platform at any given time;

[0031] express The variance of the signal-to-interference-plus-noise ratio (SIR) among all user devices at any given time;

[0032] These are the adjustable weight coefficients for each item.

[0033] Furthermore, as a preferred embodiment of the present invention, the machine learning decision module adopts an offline pre-training method during the training phase, using historical multi-scenario channel data and simulation environment to generate a large number of trajectory samples to initialize the model; during the running phase, it adopts an online incremental learning or few-sample fine-tuning mechanism to continuously optimize the model parameters according to the channel characteristics of the current actual deployment area.

[0034] Furthermore, as a preferred embodiment of the present invention, there are at least two cooperative modes between the unmanned aerial vehicle platform and the intelligent reflective surface, including:

[0035] In the first collaborative mode, the intelligent reflective surface is fixedly deployed on the ground or on a high building, and the UAV acts as an aerial mobile relay to provide dynamic line-of-sight supplementation.

[0036] In the second collaborative mode, the intelligent reflective surface is mounted on the UAV platform to form an aerial mobile intelligent reflective unit, in which the UAV simultaneously undertakes the dual roles of signal forwarding and reflection control.

[0037] The adaptive switching module for the cooperation mode automatically selects or mixes the first cooperation mode and the second cooperation mode based on the real-time link congestion level and energy consumption constraints.

[0038] Furthermore, in a preferred embodiment of the present invention, the input feature vector of the machine learning decision module includes, but is not limited to:

[0039] Real-time 3D position coordinates, velocity vector, and attitude angles of the UAV;

[0040] Geographic location, service type priority, and instantaneous channel gain estimation for multi-user equipment;

[0041] The current phase state of each reflective element of the intelligent reflective surface;

[0042] Estimation of the location and intensity of environmental disturbance sources;

[0043] The multi-source heterogeneous environment information refers to the collective term for various types of information obtained from different physical sources, with different data structures and physical meanings, used to describe the overall picture of the communication environment;

[0044] Preferably, the attention mechanism is a multi-head attention mechanism, and the graph neural network is a graph attention network.

[0045] Furthermore, as a preferred embodiment of the present invention, the deep reinforcement learning network adopts an Actor-Critic architecture, wherein the Critic network evaluates the long-term reward of the state-action pair, and the Actor network outputs a continuous or discrete modulating action space.

[0046] The action space includes incremental adjustments to the drone's speed and heading, and discrete quantization steps or continuous phase values ​​for the phase displacement of the intelligent reflector.

[0047] Furthermore, as a preferred embodiment of the present invention, the system further includes a security enhancement submodule, which is used to superimpose physical layer security constraints on the control strategy generated by the machine learning decision module, and preferentially generate reflection direction and power allocation schemes that can suppress known or potential eavesdropping nodes.

[0048] Furthermore, as a preferred embodiment of the present invention, the system supports a multi-UAV collaborative working mode, in which multiple UAVs carry or collaboratively control multiple intelligent reflective surfaces, and the machine learning decision module adopts a multi-agent reinforcement learning framework or federated learning mechanism to achieve distributed collaborative decision-making.

[0049] A second aspect of the present invention provides a communication control method for implementing the aforementioned system method, comprising the following steps:

[0050] Collect real-time environmental information from multiple sources;

[0051] The multi-modal perception fusion module performs spatiotemporal feature fusion on the multi-source real-time environmental information to generate a high-dimensional state representation vector;

[0052] The high-dimensional state representation vector is input into a pre-trained and online optimized machine learning decision model;

[0053] The model outputs commands for adjusting the UAV's flight attitude and commands for controlling the phase shift of the intelligent reflector.

[0054] The drone trajectory and the reflection coefficient of the intelligent reflective surface are adjusted synchronously according to the instructions to achieve dynamic enhancement of the communication link.

[0055] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described communication control method.

[0056] Compared with the prior art, the beneficial effects of the present invention and its preferred embodiments are reflected in, but are not limited to, the following aspects:

[0057] 1. Improved accuracy and real-time performance of decision-making: By constructing a multimodal perception fusion module, multi-source heterogeneous information such as UAV status, channel, and user needs are deeply fused in time and space to generate a more accurate environmental state representation. Combined with an end-to-end machine learning decision-making module, it realizes direct mapping from raw perception to joint control commands, improving the accuracy and response speed of UAV trajectory and intelligent reflector phase collaborative optimization in highly dynamic environments.

[0058] 2. Enhanced system environmental adaptability and robustness: By introducing online incremental learning or few-sample fine-tuning mechanisms, the deployed model can continuously optimize its parameters based on real-time feedback data in the actual environment, quickly adapt to changes in channel characteristics in different geographical scenarios (such as cities, suburbs, and mountains), effectively solve the problem of performance degradation after deployment of traditional offline training models, and improve the long-term reliability of the system.

[0059] 3. Optimized system energy efficiency and service quality: Through the adaptive switching module of collaborative mode, the system can intelligently switch between multiple modes such as direct connection, reflection enhancement, and relay based on dynamic factors such as link quality and remaining energy. This ability to dynamically select the optimal working mode effectively avoids unnecessary energy consumption and communication interruption, and maximizes the system's energy efficiency while ensuring service quality.

[0060] 4. Expanded system functionality and security: By supporting the collaborative work of multiple drones and multiple intelligent reflectors, the system can cover a wider area and meet more complex communication needs. At the same time, the introduction of the physical layer security enhancement submodule enables the system to actively avoid the risk of eavesdropping while optimizing communication performance, thereby improving the security of the communication process.

[0061] 5. By introducing a cascaded channel representation model and a standardized Markov decision process model, this invention provides dual model support for the system at both the physical and algorithmic layers, overcoming the shortcomings of traditional schemes that are general, lack basis, and are not quantifiable, and further improving the completeness and feasibility of the technical solution. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the overall architecture of the UAV communication system based on machine learning and intelligent reflective surface assistance in Embodiment 1 of the present invention;

[0063] Figure 2 for Figure 1 A schematic diagram of the structure of the machine learning decision module and its internal sub-modules;

[0064] Figure 3 This is an overall flowchart of the communication control method in Embodiment 2 of the present invention;

[0065] Figure 4 This is a flowchart illustrating the decision-making process for adaptive switching of collaboration modes in Embodiment 1 of the present invention. Detailed Implementation

[0066] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] It should be noted that in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] It should be noted that the core processing unit of the present invention (including a multimodal perception fusion module, a machine learning decision module, and a collaborative mode adaptive switching module) can be deployed on a cloud server, a ground base station, or an edge computing unit carried by a drone platform. These different physical deployment methods are all equivalent transformations of the present invention and should all fall within the protection scope of the present invention.

[0069] Example 1

[0070] Please see Figure 1 This embodiment provides a drone communication system based on machine learning and intelligent reflective surface assistance. The system includes: a drone platform, a ground base station, an intelligent reflective surface, at least one user device, and a core processing unit (which can be deployed on a cloud server, the ground base station side, or an edge computing unit carried by the drone platform).

[0071] The unmanned aerial vehicle (UAV) platform is equipped with a communication module and a flight control module. The communication module is used to communicate wirelessly with ground base stations and user equipment, and the flight control module is used to control the UAV's flight attitude, speed, and trajectory according to the received instructions.

[0072] The intelligent reflective surface is equipped with multiple controllable reflective units. Each reflective unit can independently adjust the phase and / or amplitude of its incident wireless signal, thereby changing the propagation direction and intensity of the signal. The intelligent reflective surface can communicate with the core processing unit via wired or wireless means to receive phase adjustment commands.

[0073] The core processing unit is the core of the system of the present invention, and its internal logical functional modules include: a multimodal perception fusion module, a machine learning decision module, and a collaborative mode adaptive switching module.

[0074] The multimodal sensing fusion module is used to collect and fuse multi-source heterogeneous environmental information in real time, including but not limited to:

[0075] Unmanned aerial vehicle (UAV) platform status data: from the UAV platform's inertial measurement unit (IMU), global positioning system (GPS), etc., including three-dimensional position coordinates. Velocity vector Attitude angles (pitch angle, yaw angle, roll angle), etc.

[0076] Channel state information: from communication modules of ground base stations and / or UAV platforms, including cascaded channel gain estimation, signal-to-interference-plus-noise ratio (SINR), etc.

[0077] User equipment requirement information: from each user equipment, including its geographical location, service type (such as real-time video, voice, data) and service quality requirements (such as minimum rate requirements, maximum latency tolerance).

[0078] Environmental interference information: Location and intensity estimation of environmental interference sources obtained through spectrum sensing or a preset database.

[0079] This module first performs time synchronization and preprocessing on the aforementioned data. Then, it uses an attention mechanism or graph neural network to extract and fuse spatiotemporal features from the multimodal data, generating a high-dimensional state representation vector that comprehensively reflects the environmental state. .

[0080] The system uses a channel modeling unit to perform cascaded channel characterization of the wireless propagation links between base stations, smart reflectors, UAVs, and user equipment, providing a physical layer model basis for high-dimensional state characterization vectors, making the state characterization more reasonable and deterministic.

[0081] The term "multi-source heterogeneous environment information" refers to a collective term for various types of information obtained from different physical sources, with different data structures and physical meanings, used to describe the overall picture of the communication environment.

[0082] The machine learning decision module receives the high-dimensional state representation vector. Based on a pre-trained deep reinforcement learning model that supports online fine-tuning, a joint control policy is generated end-to-end. Please refer to [reference needed]. Figure 2 This module may specifically include a strategy generation submodule. It should be noted that in this embodiment, the environmental perception function has been implemented by the multimodal perception fusion module, and the machine learning decision module directly reuses its output high-dimensional state representation vector, without having to repeatedly set up the environmental perception submodule.

[0083] Specifically, the strategy generation submodule employs a deep neural network, with the input layer receiving the high-dimensional state representation vector. The output layer corresponds to the attitude adjustment parameters of the UAV and the phase control parameters of each unit of the intelligent reflector, respectively, and the joint strategy is generated through end-to-end training.

[0084] The machine learning decision module learns and outputs policies based on the Markov decision process model. It standardizes and defines the state space, action space, and reward function, so that end-to-end decision-making has a clear model basis and execution logic, avoiding vague decision-making.

[0085] The policy generation submodule employs a deep reinforcement learning network, such as the Actor-Critic architecture, where the Actor network receives the state. Output joint action This action also includes commands to adjust the flight attitude of the drone platform (such as three-dimensional position increments). ) and commands for controlling the reflection phase shift matrix of the smart reflector (e.g., the phase adjustment amount for each reflector unit). , Specifically, joint action vectors It can be formalized as ,in express Joint action vector at time step, This represents the three-dimensional increment of the drone's position. This represents the increment of the drone's attitude angles (roll, pitch, yaw). This represents the diagonal matrix of reflection coefficients of the intelligent reflective surface. Its vectorized form includes the phase adjustment amount of each reflection unit, and each element The range is (Continuous control) or discrete quantization into a finite number of phase values; to ensure the continuity and stability of the operation, the network output layer can adopt... After the activation function is multiplied by an appropriate range scaling, or a Gaussian policy is used to output the action mean and diagonal covariance matrix, the Critic network then evaluates the current state. Take action below value This module guides the policy gradient updates of the Actor network and can be trained using deep reinforcement learning algorithms. The reward function... The design employs a multi-objective weighted approach to guide the model in learning and optimizing the overall system performance.

[0086] ;

[0087] in, express Instant reward value at any moment express All moments The sum of the instantaneous data transmission rates of each user device express The estimated probability of system link interruption at any given time can be estimated based on real-time channel quality and a preset interruption threshold. express Energy consumption estimates for drone platforms at any given time. express The variance of the signal-to-interference-plus-noise ratio (SIR) among all user devices at any given time. These are the adjustable weight coefficients for each item;

[0088] The throughput term is used to maximize the total system transmission rate, the interruption probability term is used to improve link reliability, the drone energy consumption term is used to extend the operation time, and the fairness term is used to balance the service quality among multiple users.

[0089] The adaptive switching module for the collaboration mode is used to dynamically select the optimal UAV-intelligent reflector collaboration mode according to the real-time environment. This embodiment presets two collaboration modes (such as...). Figure 4 (as shown)

[0090] First collaborative mode (smart reflector fixed deployment auxiliary mode): The smart reflector is fixedly deployed on the ground or on a high building, and the UAV platform acts as an aerial mobile relay, dynamically adjusting its position to form the best collaborative link with the smart reflector.

[0091] The second collaborative mode (smart reflector airborne mode): The smart reflector is mounted on the UAV platform to form a mobile aerial smart reflector. The UAV simultaneously plays the dual role of signal relay and phase control of the smart reflector.

[0092] Based on real-time information such as the line-of-sight probability of the current link, the remaining energy of the UAV, the user service priority, and the degree of link congestion, this module evaluates the overall effectiveness of each cooperation mode. When the overall effectiveness of the airborne mode is better than that of the ground fixed mode by more than a preset threshold, the mode switching process is triggered. Alternatively, a soft switching method can be used, that is, the phase weights of the two modes are used in combination to achieve a smooth transition.

[0093] Through the collaborative work of the three core modules mentioned above, this system can achieve real-time joint optimization of UAV trajectory and intelligent reflector phase in highly dynamic environments, and has the ability to adaptively switch cooperative modes, thereby significantly improving the performance and robustness of the communication system.

[0094] Example 2

[0095] Please see Figure 3 This embodiment provides a communication control method based on machine learning and intelligent reflective surface assistance, which can be applied to the system described in Embodiment 1. The method includes the following steps:

[0096] S1: System initialization, loading offline pre-trained machine learning decision model parameters. The multi-modal perception fusion module collects and fuses multi-source real-time environmental information from the initial moment to generate a high-dimensional state representation vector. .

[0097] S2: Transfer the state representation vector The input to the strategy generation submodule of the machine learning decision module is used to infer the current optimal joint control strategy, including the UAV flight attitude adjustment command and the intelligent reflector phase shift control command.

[0098] S3: The above instructions are sent to the UAV platform and the intelligent reflector for execution. The UAV adjusts its flight trajectory and attitude, and the intelligent reflector updates the phase shift configuration of each reflector unit in sync.

[0099] S4: After execution, collect actual communication performance feedback (throughput, interruption status, energy consumption, etc.) and calculate the instant reward value. And obtain the state at the next moment. .

[0100] S5: Set the current experience tuple Store in the experience replay buffer.

[0101] S6: The adaptive switching module for collaborative modes evaluates the applicability of the current collaborative mode based on information such as real-time link quality, remaining drone energy, and user priority. If the preset switching conditions are met, the mode switching process is triggered to update the system's working mode.

[0102] S7: Periodically sample small batches of experience tuples from the experience replay buffer, and use the policy gradient method to perform online incremental learning or fine-tuning of the machine learning decision model, so that the model parameters can gradually adapt to the channel characteristics of the current real environment. The online learning here can be achieved by techniques such as experience replay, policy gradient update or few-sample fine-tuning based on meta-learning.

[0103] S8: Repeat steps S2 to S7 until the communication task is completed or the preset communication duration is reached.

[0104] It should be noted that the mode switching decision in step S6 and the online model fine-tuning in step S7 are two parallel processes, executed by independent sub-modules respectively, to jointly ensure the optimal operating state of the system.

[0105] Example 3

[0106] This embodiment, based on embodiment one or two, further adds security enhancement functions. The system also includes a security enhancement submodule, which can be integrated into the decision-making process in various ways:

[0107] Reward function shaping: Add physical layer security-related terms to the reward function of the machine learning decision module. For example, add confidentiality capacity as a reward term to improve physical layer security, or add negative values ​​of confidentiality interruption probability as a penalty term. In this way, the model will naturally tend to choose strategies that can improve security during the optimization process.

[0108] Action space constraints: After the strategy generation submodule outputs the intelligent reflector phase adjustment command, the security enhancement submodule performs post-processing on it. For example, if the location of a potential eavesdropper is known, the phase matrix can be fine-tuned through optimization algorithms to form a null in the direction of the eavesdropper, thereby reducing the strength of the received signal.

[0109] Artificial noise assistance: The security enhancement submodule can control the UAV platform to superimpose artificial noise orthogonal to the legitimate user's channel when transmitting signals. This noise has no impact on the legitimate user's reception, but can effectively interfere with the eavesdropper's decoding.

[0110] By employing the above methods, this system can effectively enhance the physical layer security of the communication process and prevent information leakage while pursuing optimal communication performance.

[0111] Example 4

[0112] This embodiment extends the first embodiment to a multi-drone collaborative scenario. The system supports the collaborative work of multiple drone platforms, and each drone can carry or collaboratively control one or more intelligent reflective surfaces.

[0113] In this scenario, the machine learning decision-making module adopts a multi-agent reinforcement learning framework, such as a multi-agent reinforcement learning algorithm. Each UAV acts as an agent, possessing its own local observation information (such as its own position and local channel state), and shares some information (such as global user distribution and interference source location) with other agents through limited communication links. The reward function of each agent can be designed as centralized or hybrid, aiming to jointly optimize the global objectives at the system level, such as maximizing the area coverage or minimizing the total system interruption probability.

[0114] Alternatively, the machine learning decision module can also adopt a federated learning mechanism, where each UAV uses local data to train the model and only uploads the model gradient to the central server for aggregation, thereby achieving distributed collaborative decision-making while protecting data privacy.

[0115] This implementation method is particularly suitable for complex scenarios that require multi-platform collaboration, such as large-scale emergency communication restoration and temporary network coverage for large-scale events.

[0116] In summary, this invention establishes a complete, closed-loop, and self-optimizing UAV-intelligent reflector cooperative communication solution by constructing three core modules: multimodal perception fusion, end-to-end machine learning decision-making, and adaptive cooperative mode switching. This system effectively addresses communication challenges in highly dynamic and complex environments, demonstrating significant advantages in improving communication quality, system energy efficiency, and environmental adaptability.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A UAV communication system based on machine learning and intelligent reflective surface assistance, characterized in that, include: Unmanned aerial vehicle platform, ground base station, smart reflector and at least one user device; The unmanned aerial vehicle platform is equipped with a communication module and a flight control module; The intelligent reflective surface is equipped with multiple controllable reflective units for phase and / or amplitude modulation of incident wireless signals. It also includes a channel modeling unit, which is used to construct a cascaded channel representation model of base station-smart reflector-UAV-user, providing physical layer model support for the system; The multimodal perception fusion module is used to collect and fuse multi-source heterogeneous environmental information, extract spatiotemporal features through attention mechanisms or graph neural networks, and generate high-dimensional state representation vectors. The machine learning decision module is used to receive the high-dimensional state representation vector and generate a joint control strategy end-to-end based on the Markov decision process model and the deep neural network model. The joint control strategy includes flight attitude adjustment instructions for the UAV platform and control instructions for the reflection phase shift matrix of the intelligent reflector. The machine learning decision module is also used to generate a theoretical representation of the system state of link performance, energy efficiency and decision adaptability. The adaptive switching module for collaboration mode is used to dynamically select or mix at least two preset UAV-intelligent reflector collaboration modes based on real-time environmental information. The collaboration modes include a first collaboration mode and a second collaboration mode, wherein the first collaboration mode is a fixed deployment auxiliary mode for intelligent reflectors and the second collaboration mode is an airborne mode for intelligent reflectors. The high-dimensional state representation vector generated by the multimodal perception fusion module serves as the input basis for both the machine learning decision module and the collaborative mode adaptive switching module, thereby achieving closed-loop optimization of perception, modeling, decision-making, and reconstruction.

2. The system according to claim 1, characterized in that, The machine learning decision module includes: The strategy generation submodule uses a deep reinforcement learning network or a multi-task deep learning network to output the flight attitude adjustment command and the reflection phase shift matrix control command in real time according to the preset reward function. The reward function includes at least a weighted factor for multiple objectives, such as communication throughput, negative link interruption probability, drone energy consumption, and service fairness.

3. The system according to claim 1, characterized in that, The machine learning decision module adopts an offline pre-training method during the training phase, using historical multi-scenario channel data and simulation environment to generate a large number of trajectory samples to initialize the model; during the running phase, it adopts an online incremental learning or few-sample fine-tuning mechanism to continuously optimize the model parameters according to the channel characteristics of the current actual deployment area.

4. The system according to claim 1, characterized in that, The collaboration modes include: In the first collaborative mode, the intelligent reflective surface is fixedly deployed on the ground or on a high building, and the UAV acts as an aerial mobile relay to provide dynamic line-of-sight supplementation. In the second collaborative mode, the intelligent reflective surface is mounted on the UAV platform to form an aerial mobile intelligent reflective unit, in which the UAV simultaneously undertakes the dual roles of signal forwarding and reflection control. The adaptive switching module for the cooperation mode automatically selects or mixes the first cooperation mode and the second cooperation mode based on the real-time link congestion level and energy consumption constraints.

5. The system according to claim 1, characterized in that, The multi-source heterogeneous environmental information constitutes the input feature vector, including but not limited to: Real-time 3D position coordinates, velocity vector, and attitude angles of the UAV; Geographic location, service type priority, and instantaneous channel gain estimation for multi-user equipment; The current phase state of each reflective element of the intelligent reflective surface; Estimation of the location and intensity of environmental disturbance sources; The multimodal perception fusion module uses an attention mechanism or a graph neural network to extract and fuse spatiotemporal features from the input feature vector to generate the high-dimensional state representation vector.

6. The system according to claim 2, characterized in that, The deep reinforcement learning network adopts an Actor-Critic architecture, in which the Critic network evaluates the long-term reward of the state-action pair, the Actor network outputs a continuous or discrete modulating action space, and the output of the Critic network is used to guide the policy update of the Actor network. The action space includes incremental adjustments to the drone's speed and heading, and discrete quantization steps or continuous phase values ​​for the phase displacement of the intelligent reflector.

7. The system according to claim 1, characterized in that, It also includes a security enhancement submodule, which is used to superimpose physical layer security constraints on the control strategy generated by the machine learning decision module, and preferentially generate reflection direction and power allocation schemes that can suppress known or potential eavesdropping nodes.

8. The system according to claim 1, characterized in that, The system supports a multi-UAV collaborative working mode, with multiple UAVs carrying or cooperating to control multiple intelligent reflective surfaces. The machine learning decision-making module uses a multi-agent reinforcement learning framework or federated learning mechanism to achieve distributed collaborative decision-making.

9. A communication control method for implementing the system of claim 1, characterized in that, Includes the following steps: Collect real-time environmental information from multiple sources; Based on the multi-source real-time environmental information, dynamic selection or mixed use is performed between at least two preset UAV-intelligent reflector cooperation modes; The multi-modal perception fusion module performs spatiotemporal feature fusion on the multi-source real-time environmental information to generate a high-dimensional state representation vector; The high-dimensional state representation vector is input into a pre-trained and online optimized machine learning decision model; The model outputs commands for adjusting the UAV's flight attitude and commands for controlling the phase shift of the intelligent reflector. The drone trajectory and the reflection coefficient of the intelligent reflective surface are adjusted synchronously according to the instructions to achieve dynamic enhancement of the communication link.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the communication control method of claim 9.

Images