A method for unmanned aerial vehicle communication capacity maximization beam adaptive prediction and active compensation for non-stationary turbulent environment

Abstract

The application discloses a UAV communication capacity maximization beam adaptive prediction and active compensation method for a non-stationary turbulent environment, and comprises the following steps: collecting multi-modal original data, aligning the multi-modal original data, obtaining a multi-modal joint state vector of the aligned multi-modal original data, inputting the multi-modal joint state vector into a prediction model based on flow state perception and capacity maximization to obtain an optimal beam correction vector for maximum recovery of damaged capacity, generating an anti-disturbance phase compensation matrix based on the optimal beam correction vector for maximum recovery of damaged capacity, and compensating and correcting a UAV millimeter wave communication beam based on the anti-disturbance phase compensation matrix; and monitoring the corrected UAV millimeter wave communication beam in real time by using a non-smooth environment mutation model, and dynamically adjusting parameters of the model based on a monitoring result. Through fusion of physical mechanism perception, performance-oriented prediction and mutation adaptive evolution, the application effectively solves the problems of feedback lag, model mismatch and incapability of coping with a mutation environment of a traditional geometric beam tracking under complex weather conditions.

Description

Technical Field

[0001] This invention belongs to the fields of wireless communication, intelligent signal processing and deep learning application technology, and specifically relates to a method for maximizing the communication capacity of unmanned aerial vehicles (UAVs) through adaptive beam prediction and active compensation in non-stationary turbulent environments. Background Technology

[0002] With the development of 6G network architecture and the low-altitude economy, drones widely adopt massive MIMO and narrow beam technology to combat propagation loss. However, high-frequency micro-flickering of the aircraft and channel amplitude flicker caused by low-altitude atmospheric turbulence severely restrict link reliability. Existing beam control technologies still have some problems: traditional pilot tracking mechanisms are limited by feedback time delay and cannot respond to high-frequency turbulence in real time; geometric alignment strategies ignore capacity optimization in strong flicker environments and suffer from gradient saturation in high signal-to-noise ratio regions, making it difficult to finely improve throughput; and deep learning models based on offline training assume a stable environment and cannot cope with "concept drift" caused by wind shear or building obstruction, which is prone to catastrophic forgetting. Therefore, there is an urgent need for a beam intelligent control method that integrates physical flow perception, takes capacity maximization as its direct goal, and has the ability to monitor and adaptively evolve non-smooth abrupt environments online. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention proposes an adaptive beam prediction and active compensation method for maximizing the communication capacity of unmanned aerial vehicles (UAVs) in non-stationary turbulent environments. This method includes: constructing a multimodal heterogeneous data synchronous sensing mechanism and collecting multimodal raw data, aligning the multimodal raw data; obtaining the multimodal joint state vector of the aligned multimodal raw data; constructing a prediction model based on flow perception and capacity maximization, inputting the multimodal joint state vector into the prediction model to obtain the optimal beam correction vector that maximizes the recovery of damaged capacity; generating an anti-disturbance phase compensation matrix based on the optimal beam correction vector that maximizes the recovery of damaged capacity, and compensating and correcting the UAV millimeter-wave communication beam based on the anti-disturbance phase compensation matrix; and using a non-smooth environment mutation model to monitor the corrected UAV millimeter-wave communication beam in real time, and dynamically adjusting the model parameters based on the monitoring results.

[0004] The beneficial effects of this invention are:

[0005] This invention effectively solves the problems of feedback lag, model mismatch, and inability to cope with sudden environmental changes in traditional geometric beam tracking under complex weather conditions by integrating physical mechanism perception, performance-oriented prediction, and mutation adaptive evolution, and significantly improves the robustness and effective throughput of UAV communication systems. Attached Figure Description

[0006] Figure 1This is a system model diagram of the present invention;

[0007] Figure 2 This is the overall flowchart of the present invention. Detailed Implementation

[0008] 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.

[0009] An adaptive beam prediction and active compensation method for maximizing communication capacity of unmanned aerial vehicles (UAVs) in non-stationary turbulent environments, such as... Figure 1 and Figure 2 As shown, the method includes: constructing a multimodal heterogeneous data synchronous sensing mechanism and collecting multimodal raw data, and aligning the multimodal raw data; obtaining the multimodal joint state vector of the aligned multimodal raw data; constructing a prediction model based on flow perception and capacity maximization, inputting the multimodal joint state vector into the prediction model based on flow perception and capacity maximization to obtain the optimal beam correction vector that maximizes the recovery of damaged capacity; generating an anti-disturbance phase compensation matrix based on the optimal beam correction vector that maximizes the recovery of damaged capacity, and compensating and correcting the UAV millimeter-wave communication beam based on the anti-disturbance phase compensation matrix; and using a non-smooth environmental mutation model to monitor the corrected UAV millimeter-wave communication beam in real time, and dynamically adjusting the model parameters based on the monitoring results.

[0010] In this embodiment, a robust beam prediction and capacity maximization compensation method for UAV millimeter-wave communication based on turbulence statistical feature perception includes:

[0011] S1: Construct a multimodal heterogeneous data synchronization and perception mechanism.

[0012] During communication missions, the system concurrently activates physical motion sensing channels and radio frequency (RF) channel sensing channels within the onboard computing unit to construct the raw dataset. The physical sensing channel utilizes an inertial measurement unit at a high sampling rate. Real-time acquisition of dynamic response data of the aircraft under turbulent wind fields captures millisecond-level micro-vibrations. The radio frequency channel sensing channel, through a millimeter-wave radio frequency front-end and a baseband processor, acquires data at frame rates. Obtain the current channel state information and beam reference signal received power. Given the heterogeneity of the two data sources, assign a uniform microsecond-level timestamp to each frame of IMU and CSI data, and store the aligned data in a circular buffer to provide rigorous timing support for subsequent turbulence feature extraction. The acquired raw inertial data vector... With channel state vector The definition is as follows:

[0013]

[0014]

[0015] in, The specific force component measured by a triaxial linear accelerometer. The angular velocity component of the machine body measured by a three-axis gyroscope; The number of receiving antenna elements. and The first The channel amplitude response and phase response of each array element on the current subcarrier.

[0016] S2: Deeply explore the joint features of "turbulence statistics-communication capacity".

[0017] After obtaining the time-aligned multimodal raw data from S1, this step extracts turbulence statistical features that characterize environmental physical properties and capacity evolution features that characterize link quality trends in sequence, based on the physical properties of the acquired data, and completes the fusion of multidimensional features.

[0018] Furthermore, the step S21 for extracting turbulent statistical features and determining the flow regime in S2 is as follows:

[0019] S21: Extract turbulent statistical features and determine the flow regime.

[0020] This step aims to quantify the severity of the current atmospheric environment and provide environmental information for the prediction model. The system calculates statistical indicators from two physical dimensions: mechanical dynamics and electromagnetic propagation, and classifies the current environmental flow regime accordingly.

[0021] S211: This step takes arrive This period is used as a sliding window to retrieve... Using acceleration data within a sliding window, the variance of mechanical turbulence is calculated from a set of data points. This index is used to characterize the severity of airflow disturbance on the physical structure of a drone. The formula for calculating this index is:

[0022]

[0023] in, The length of the sliding window. The three-axis acceleration vector at a historical moment. This represents the average acceleration magnitude within the window. The larger this value, the less stable the organism.

[0024] S212: This step uses the CSI channel amplitude data from the sliding window in the previous step to calculate the channel scintillation index. This index is used to characterize the effect of random fluctuations in atmospheric refractive index on the amplitude of electromagnetic waves. The formula for calculating this index is:

[0025]

[0026] in, This represents the ensemble averaging operation within the sliding window, and this indicator directly reflects the dispersion of the channel amplitude.

[0027] S213: Based on the two physical parameters of steps S211 and S212, the system introduces a flow regime flag. The current flight environment is classified in real time. Specifically, when the mechanical turbulence variance... With channel scintillation index When all indicators are below the preset safety threshold, the system determines the current environment to be in a steady state (Level 0). At this time, the flight environment is stable, and the communication link is mainly affected by Gaussian white noise. When only one indicator is detected to exceed the corresponding threshold, the system determines it to be in a weak turbulent state (Level 1), indicating that there is a single-dimensional interference, that is, only the aircraft is affected by gusts of wind or only the channel is affected by refractive index flicker. When both indicators are detected to exceed the threshold, the system determines it to be in a strong turbulent state (Level 2), indicating that the UAV is in a complex strong convective turbulence, the communication link is very easy to be interrupted, and the highest level of prediction compensation mechanism needs to be activated.

[0028] S22: Extract the absolute performance status features of the communication domain.

[0029] This step quantifies the absolute performance loss caused by the turbulence level determined by S21 to the current communication link. By calculating the gap between the current index and the ideal index, the health of the link is reflected intuitively, thus providing a clear optimization gain space for the prediction model.

[0030] S221: Calculate instantaneous capacity utilization rate This metric is used to visually characterize the degree to which the current channel capacity is maintained relative to the theoretical peak. It represents the instantaneous channel capacity at the current moment. Compared with the theoretical maximum capacity under the current modulation and coding strategy The comparison is performed, and the calculation formula is as follows:

[0031]

[0032] in, This refers to the system bandwidth. The range of values ​​is The closer this value is to 1, the closer the link is to a perfect state; the lower the value, the more severe the capacity collapse caused by turbulence, and the greater the capacity gap that the model needs to compensate for.

[0033] S222: To overcome the computational limitations of S221 To address gradient saturation at high signal-to-noise ratios and ensure physical layer alignment accuracy, this step calculates the beam gain loss. This is used to directly quantify the received energy loss caused by machine jitter. The system uses the theoretical value of line-of-sight transmission from the base station. Used as a reference received power, and compared with the current beam reference signal received power. Perform difference calculations:

[0034]

[0035] This indicator directly reflects the energy loss caused by the main lobe of the beam deviating from the receiver. The larger the value, the greater the beam pointing error, and the greater the phase compensation required.

[0036] S23: Construct a multimodal joint state vector. Specifically: extract physical flow characteristics representing the physical properties of the current environment based on mechanical turbulence variance, channel scintillation index, and their corresponding flow state flags; extract the three-axis acceleration, three-axis angular velocity, and attitude angles of the aircraft at the current moment using real-time data collected by the inertial measurement unit, and construct motion characteristics representing the instantaneous motion state of the aircraft; combine the communication absolute performance characteristics composed of instantaneous capacity utilization and beam gain loss to obtain the physical flow characteristics, communication absolute performance characteristics, and the current instantaneous motion state of the aircraft.

[0037] The extracted physical flow characteristics and absolute communication performance characteristics are then concatenated with the current instantaneous motion state of the organism to construct a time-space model. The joint eigenvector of turbulence-performance The vector is defined as:

[0038]

[0039] This vector, as the input tensor, will be fed into the subsequent Attention-LSTM model to predict the optimal beam correction that maximizes future communication capacity.

[0040] S3: Construct a prediction model based on flow perception and capacity maximization.

[0041] This step constructs a deep learning model capable of understanding the nonlinear mapping relationship between the physical environment and communication performance. It uses the "turbulence-performance" joint feature vector sequence constructed by S2 as input, captures the temporal evolution of body jitter and channel fading through a long short-term memory network, and introduces a flow-state-based attention mechanism to enhance the model's feature extraction capability in strong turbulent environments. Furthermore, this invention proposes maximizing communication link capacity as the direct optimization objective, thereby ensuring that the beam correction strategy output by the model can restore the damaged spectral efficiency to the greatest extent possible.

[0042] Furthermore, the Attention-LSTM network structure S31 is constructed in S3 as follows:

[0043] S31: Construct an Attention-LSTM network structure for flow perception.

[0044] To enable the prediction model to automatically adjust its focus based on the current atmospheric flow, this invention constructs an Attention-LSTM network that integrates flow-gated control. The model first utilizes multiple stacked LSTM units to extract the input time series. The deep temporal features in the data yield the hidden state sequence. Based on this, the system introduces flow regime flags. The query vector for the attention mechanism is used to calculate the attention weight at each historical moment. The calculation formula is:

[0045]

[0046] in , , , These are trainable parameters. The mechanism aims to achieve adaptive attention to the environment: when... When the current state is strongly turbulent, the network will automatically assign those mechanical turbulence variances. or capacity drop gradient Historical moments experiencing drastic fluctuations are given higher weights, forcing the model to focus on learning the impact patterns of sudden disturbances on the link; while in steady-state conditions, historical information is given more even attention. Finally, a context vector is obtained through weighted summation, and future moments are output via a fully connected layer. Optimal beam correction vector .

[0047] The hidden state sequence is weighted and summed based on the obtained attention weights to obtain a context vector that integrates flow perception information. The context vector is then input into a fully connected layer and decoded using a capacity maximization criterion to output the optimal beam correction vector that can restore the damaged communication capacity to the greatest extent in the next time step.

[0048] S32: Define the communication capacity maximization loss function.

[0049] To address the physical challenge that geometric alignment does not equate to optimal performance, particularly to cope with the nonlinear fading of beam gain caused by channel flicker, this invention designs a negative capacity loss function directly oriented towards communication metrics. During offline model training and online fine-tuning, the system incorporates the predicted beam correction amount... Substituting these parameters into the channel capacity calculation formula aims to find a set of network parameters that, under this correction, maximize the approximation of the theoretical capacity upper limit. Simultaneously, a regularization term is introduced to ensure the smoothness of beam control and prevent model overfitting. Loss function The definition is as follows:

[0050]

[0051] in, For training batch size, This represents the theoretical maximum communication capacity of the current system under ideal alignment and a stable channel condition. For transmission power, Let be the directional gain function of the transmitting antenna array. Given the current physical line-of-sight angle, This represents the corresponding channel fading coefficient. For noise power, This is a regularization penalty term, which includes corrections for the predicted beam. A penalty is applied to suppress ineffective high-frequency micro-deflections of the UAV in non-stationary environments. The gradient descent direction of this loss function no longer approximates the true physical angle, but rather the equivalent angle that yields the maximum signal-to-noise ratio. This means that even if there is a slight deviation between the predicted angle and the geometric center, if this deviation can utilize the multipath effect to obtain a higher synthetic gain, the model will still be considered correct, thus achieving true communication performance-oriented optimization.

[0052] S4: Active beam precoding and RF front-end phase compensation.

[0053] Based on the optimal beam correction vector that maximizes communication capacity output in step S3, the system enters the physical layer execution phase, calculates the anti-disturbance phase compensation matrix, and updates the RF front-end configuration. This step utilizes the predictive nature of the information to perform reverse pre-deflection of the transmit beam at the synchronization moment when jitter actually occurs in the device or channel flicker occurs, thereby offsetting the gain loss caused by environmental interference and ensuring that the instantaneous signal-to-noise ratio and spectral efficiency of the receiver are maintained at the optimal level.

[0054] Furthermore, the anti-disturbance target beam pointing S41 is calculated in S4 as follows:

[0055] S41: Calculate the beam pointing of the anti-disturbance target.

[0056] First, based on the current reference beam pointing and the optimal correction value predicted by the model, the ideal target pointing at the next moment is calculated. Let the physical line-of-sight pointing of the UAV antenna array at the current moment be... (Corresponding to azimuth and elevation angles respectively). Combined with the correction vector based on the capacity maximization criterion output from the model in S3. Calculate the target beam pointing for the next disturbance-resistant moment. The calculation formula is:

[0057]

[0058] in, To compensate for the gain coefficient, it is generally set to a value of 1, but in weakly turbulent conditions, it is set to a value less than 1 for smoothing filtering. This target points to... This represents the absolute spatial angle at which the main lobe of the beam remains precisely aligned with the receiving base station even after the predicted attitude shift of the aircraft.

[0059] S42: Generate the phase control matrix and perform radio frequency precoding.

[0060] After determining the target's orientation, the system needs to map it into a complex weight vector of the antenna array and send it to the hardware layer. This is for drones equipped with... Uniform planar array, the system calculates the first... antenna array elements (of which) ) Required compensation phase value Based on the principle of phased array beamforming, its calculation formula is as follows:

[0061]

[0062] in, For millimeter wave carrier wavelength, The spacing between antenna elements is denoted as , and its value is . The calculated continuous phase values After quantization to match the precision of the digital phase shifter, the final phase control codebook is formed. This codebook is then written into the hybrid pre-encoder of the RF front end via the baseband control interface. This operation is completed within the guard interval before the next frame pilot signal is transmitted, ensuring that the electromagnetic beam has already completed the energy distribution reconstruction when the physical layer turbulence disturbance arrives, thereby achieving "zero-delay" capacity compensation.

[0063] S5: Online monitoring and adaptive incremental evolution of models based on non-smooth environmental mutations.

[0064] After completing phase compensation and beamforming of the RF front-end in step S4, the system enters the closed-loop monitoring phase based on actual communication feedback. This step, by constructing a closed-loop feedback mechanism, intelligently identifies smooth evolution and non-smooth abrupt changes in environmental flow (such as sudden gusts of wind shear, rainfall attenuation, or building obstruction), and dynamically adjusts the model update strategy accordingly. This solves the catastrophic forgetting problem of offline-trained models experiencing a sharp performance decline when facing non-smooth changes in the low-altitude atmospheric environment.

[0065] Specifically, real-time monitoring of the corrected UAV millimeter-wave communication beam includes: constructing a short-time sliding window, calculating the current flow characteristic distribution through the short-time sliding window, and comparing it with the historical baseline distribution; when the detection index... Exceeding the preset mutation threshold When the system determines that the current environment has undergone a non-smooth and drastic change, it immediately triggers the emergency adaptive mode; otherwise, it maintains the normal cruise mode. For different environmental states identified, it adopts a differentiated intelligent update strategy to correct the prediction model parameters based on flow perception and capacity maximization.

[0066] The differentiated intelligent update strategy for correcting the parameters of the prediction model based on flow perception and capacity maximization includes: when the UAV is in normal cruise mode, a low-rate smooth update strategy is adopted with a small learning rate (on the order of 10⁻). 4 The value range is 1×10⁻ 5 ~1×10⁻ 4 The system uniformly samples and fine-tunes from the experience buffer pool; when the UAV triggers the emergency adaptive mode, the system switches to a high-rate aggressive update strategy, activates the priority experience replay mechanism, and increases the learning rate to the order of 10⁻² (the value ranges from 1×10⁻³ to 1×10⁻², which is 10 to 100 times the learning rate of the normal cruise mode), forcing the model parameters to converge to the new environmental flow state in a very short time.

[0067] Furthermore, the environmental concept drift detection S51 in S5 is as follows:

[0068] S51: Environmental concept drift detection.

[0069] This invention does not rely on a single prediction error, but instead determines whether "concept drift" has occurred by monitoring changes in the distribution of physical features extracted from S2. The system maintains a short-term sliding window and calculates the current flow characteristic distribution (derived from the mechanical turbulence variance) in real time. and channel scintillation index The distribution is composed of KL divergence and compared with historical baseline distributions. The system introduces KL divergence as a mutation detection operator. When the detection index... Exceeding the preset mutation threshold When the system determines that a non-smooth, drastic change has occurred in the current environment (e.g., a drone suddenly entering a building-canyon wind field from an open area), it immediately triggers the emergency adaptive mode; otherwise, it maintains the normal cruise mode. The detection criterion is defined as:

[0070]

[0071] in The feature distribution of the current window, This serves as the baseline distribution. This step ensures that the system can accurately distinguish between "normal measurement noise" and "abnormal environmental mutations."

[0072] S52: Experience Replay and Incremental Update.

[0073] For different environmental states identified by S51, the system employs a differentiated intelligent update strategy to correct the parameters of the Attention-LSTM model in S3. When the drone is in normal cruise mode, a low-rate smooth update strategy is used with a smaller learning rate. The system performs fine-tuning by uniformly sampling from the experience pool to maintain model stability. When the UAV triggers emergency adaptive mode, the system switches to a high-rate aggressive update strategy, activating a priority experience replay mechanism to replay high-value samples with extremely large prediction errors that belong to the current abrupt change in flow state. Simultaneously, the learning rate is dynamically increased to... ( This forces the weights of the neural network to converge to the new environmental flow state in a very short time. The parameter update rule is as follows:

[0074]

[0075] Among them, learning rate Environmental mutation indicators Dynamic modulation, The system prioritizes replaying samples with higher weights based on their importance. Through this intelligent evolutionary mechanism that combines fast and slow approaches, the system can avoid overfitting during stable periods and achieve rapid recovery within milliseconds during periods of mutation, truly embodying the intelligent characteristic of "adapting to the environment like a living organism."

[0076] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments, characterized in that, include: A multimodal heterogeneous data synchronization and sensing mechanism is constructed, and multimodal raw data is collected and aligned. Obtain the multimodal joint state vector of the aligned multimodal raw data; construct a prediction model based on flow perception and capacity maximization, input the multimodal joint state vector into the prediction model based on flow perception and capacity maximization, and obtain the optimal beam correction vector that maximizes the recovery of damaged capacity; An anti-disturbance phase compensation matrix is ​​generated based on the optimal beam correction vector that maximizes the recovery of damaged capacity, and the UAV millimeter-wave communication beam is compensated and corrected based on the anti-disturbance phase compensation matrix; a non-smooth environmental mutation model is used to monitor the corrected UAV millimeter-wave communication beam in real time, and the model parameters are dynamically adjusted based on the monitoring results.

2. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 1, characterized in that, The acquired multimodal raw data includes airborne inertial measurement unit (IMU) data characterizing physical disturbance sources and channel state information characterizing communication performance response; specifically, it includes: physical sensing channels utilizing IMUs at high sampling rates. Real-time acquisition of dynamic response data of the aircraft under turbulent wind fields, capturing millisecond-level micro-vibrations; the radio frequency channel sensing channel, through a millimeter-wave radio frequency front-end and a baseband processor, captures data at frame rates. Obtain the current channel state information and beam reference signal received power.

3. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 1, characterized in that, The multimodal joint state vector obtained from the aligned multimodal raw data includes: Step 1: Calculate statistical indicators from two physical dimensions: mechanical power and electromagnetic propagation, and classify the current environmental flow state based on the statistical indicators; Step 2: Calculate the absolute performance loss of the interference caused by the classified environmental flow state to the current communication link, and provide a clear optimization gain space for the prediction model by calculating the gap between the current index and the ideal index, so as to obtain the physical flow state characteristics, communication absolute performance characteristics and the current instantaneous motion state of the organism. Step 3: Concatenate the physical flow characteristics, communication absolute performance characteristics and the current instantaneous motion state of the organism to obtain a multimodal joint state vector.

4. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 3, characterized in that, The process of classifying the current environmental flow regime includes: Step 11: Construct a sliding window, take K data points, and use the acceleration data within the sliding window to calculate the variance of mechanical turbulence. ; Step 12: Calculate the channel scintillation index using the CSI channel amplitude data from the sliding window. ; Step 13: Based on mechanical turbulence variance and channel scintillation index Using flow state flags Real-time classification of the current flight environment.

5. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 4, characterized in that, The physical flow characteristics, absolute communication performance characteristics, and current instantaneous motion state of the organism are obtained, including: Step 21: Obtain the instantaneous channel capacity at the current moment. Compared with the theoretical maximum capacity under the current modulation and coding strategy And calculate the instantaneous capacity utilization rate. ; Step 22: Obtain the theoretical value of line-of-sight transmission at the base station. Used as a reference received power, and compared with the current beam reference signal received power. Differential calculations are performed to obtain the beam gain loss. ; Step 23: Extract physical flow characteristics representing the physical properties of the current environment based on mechanical turbulence variance, channel scintillation index and their corresponding flow state flags; extract the three-axis acceleration, three-axis angular velocity and attitude angle of the body at the current moment using real-time data collected by the inertial measurement unit, and construct motion characteristics representing the instantaneous motion state of the body; combine the communication absolute performance characteristics composed of instantaneous capacity utilization and beam gain loss to obtain physical flow characteristics, communication absolute performance characteristics and the current instantaneous motion state of the body.

6. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 1, characterized in that, Inputting the multimodal joint state vector into a prediction model based on flow perception and capacity maximization includes: D1. An Attention-LSTM network incorporating fluid gating was constructed; multi-layer stacked LSTM units were used to extract the input time series. The deep temporal features in the data yield the hidden state sequence. Introducing flow regime flags The query vector for the attention mechanism is used to calculate the attention weight at each historical moment. ; D2. Based on the obtained attention weights, the hidden state sequence is weighted and summed to obtain the context vector that integrates the flow perception information. The context vector is input into the fully connected layer and decoded in combination with the capacity maximization criterion to output the best beam correction vector that can restore the damaged communication capacity to the greatest extent in the next time step.

7. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 6, characterized in that, The loss function for the prediction model based on flow perception and capacity maximization during training is: ； in, For training batch size, This represents the theoretical maximum communication capacity of the current system under ideal alignment and stable channel conditions. For transmission power, Let be the directional gain function of the transmitting antenna array. Given the current physical line-of-sight angle, This represents the corresponding channel fading coefficient. For noise power, This is a regularization penalty term.

8. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 1, characterized in that, Generating the anti-disturbance phase compensation matrix includes: calculating the ideal target pointing at the next moment based on the reference beam pointing at the current moment and the optimal correction amount predicted by the model; calculating the anti-disturbance target beam pointing at the next moment based on the ideal target pointing and the correction vector based on the capacity maximization criterion; mapping the anti-disturbance target beam pointing at the next moment to a complex weight vector of the antenna array and sending it to the hardware layer; and the hardware layer converting the continuous phase values ​​based on the complex weight vector. Quantization is performed to form a phase control codebook; the phase control codebook is written into the hybrid pre-encoder of the RF front end through the baseband control interface, and the energy distribution of the electromagnetic beam is reconstructed to complete zero-delay capacity compensation.

9. The method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments according to claim 1, characterized in that, Real-time monitoring of the corrected UAV millimeter-wave communication beam includes: constructing a short-time sliding window, calculating the current flow characteristic distribution through the short-time sliding window, and comparing it with the historical benchmark distribution; when the detection index Exceeding the preset mutation threshold When the system determines that the current environment has undergone a non-smooth and drastic change, it immediately triggers the emergency adaptive mode; otherwise, it maintains the normal cruise mode. For different environmental states identified, it adopts a differentiated intelligent update strategy to correct the prediction model parameters based on flow perception and capacity maximization.

10. A method for maximizing UAV communication capacity through adaptive beam prediction and active compensation in non-stationary turbulent environments, as described in claim 9, is characterized in that... The differentiated intelligent update strategy for correcting the prediction model parameters based on flow perception and capacity maximization includes: when the UAV is in normal cruise mode, a low-rate smooth update strategy is adopted to uniformly sample and fine-tune the parameters from the experience buffer pool with the learning rate; when the UAV triggers emergency adaptive mode, the system switches to a high-rate aggressive update strategy, activates the priority experience replay mechanism, and increases the learning rate to the order of 10⁻², forcing the model parameters to converge to the new environmental flow state in a very short time.

Images