Satellite mobile communication antenna adaptive beam prediction and control method

Abstract

The application relates to a satellite mobile communication antenna adaptive beam prediction and control method, which comprises the following steps: collecting motion data of a user terminal, signal state data of a satellite antenna receiving end, satellite star-ground attitude data and ground environmental interference data, and processing to obtain a multi-source data set; based on the geometric correlation characteristics, gain correlation characteristics, sidelobe correlation characteristics and array element phase correlation characteristics of the multi-source data set and beam parameters, an associated matrix is constructed, the associated matrix is spliced with the multi-source data set, and a complete input feature set is obtained; a beam prediction model is constructed, and training and verification are performed; the beam prediction model output is converted into a control signal to drive the antenna array element adjustment, and based on the link signal-to-noise ratio, the bit error rate and the throughput collected by the user terminal, the model parameters are corrected, and the beam is split / combined. The application can improve the beam parameter prediction accuracy and environmental adaptability, suppress interference, optimize resource allocation, and guarantee stable and efficient satellite communication in a high dynamic complex environment.

Description

Technical Field

[0001] This application relates to the field of radio transmission technology, and in particular to an adaptive beam prediction and control method for satellite mobile communication antennas. Background Technology

[0002] Satellite mobile communication systems, as the core support for seamless global communication, play an irreplaceable role in critical scenarios such as remote areas, maritime navigation, aviation communications, and emergency rescue, carrying the voice calls, data transmission, and multimedia interaction needs of tens of millions of users daily. Beam pointing accuracy and link stability directly determine the Quality of Service (QoS), spectrum resource utilization, and user experience. Essentially, they are the result of the coupling effects of multiple systems, including relative motion between the satellite and the ground, dynamic user movement, channel environment changes such as rain attenuation / multipath effects / electromagnetic interference, and beam resource allocation. They must address multi-dimensional dynamic challenges such as real-time changes in user location, random fluctuations in channel fading, and satellite attitude disturbances. With the increasingly urgent need for "low latency, high dynamism, and wide coverage" in low-Earth orbit satellite constellations, higher requirements are placed on the dynamic response speed, prediction accuracy, and environmental adaptability of beam control.

[0003] Existing satellite antenna beam control methods mainly include traditional digital beamforming (DBF) methods, fixed beam coverage methods, precoding-based beam optimization methods, single intelligent prediction methods, and satellite-to-ground separation control methods. However, traditional DBF methods rely on real-time channel estimation feedback, with channel estimation delays reaching 20-30ms, which cannot match the rapid relative motion between low-Earth orbit satellites and the ground, resulting in beam pointing deviations ≥0.5°. Fixed beam coverage methods divide the coverage area into static beam cells, which cannot dynamically adjust resources according to user density, leading to insufficient resources in high-density areas and wasted resources in low-density areas. Precoding-based methods do not consider real-time channel fading, resulting in mismatch between the precoding matrix and the actual channel, and a bit error rate rising to 10%. -3 The above methods exhibit several drawbacks. First, the single intelligent prediction method relies solely on user location time-series data, failing to incorporate key factors such as satellite attitude and electromagnetic interference. This results in a prediction error ≥0.3° and lacks physical constraints, making it prone to interfering with neighboring users. Second, the satellite-ground separation control method calculates beam parameters from the ground station before transmitting them, leading to a transmission delay ≥50ms, inability to handle sudden interference, and a link interruption probability ≥8%. Furthermore, none of these methods adequately characterize the "dynamic-physical-resource" coupling mechanism of beam control, resulting in large beam pointing deviations, poor link stability, and low resource utilization. Consequently, they fail to meet the low-latency, high-dynamic, and wide-coverage communication requirements of low-Earth orbit satellite constellations.

[0004] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this disclosure is to provide an adaptive beam prediction and control method for satellite mobile communication antennas, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.

[0007] This application provides a method for adaptive beam prediction and control of satellite mobile communication antennas, including:

[0008] The system collects motion data from user terminals, signal status data from satellite antenna receivers, satellite attitude data, and environmental interference data from the ground, and performs time synchronization and preprocessing to obtain a multi-source dataset.

[0009] Based on the construction of geometric correlation features, gain correlation features, sidelobe correlation features, and array element phase correlation features of the multi-source dataset and beam parameters, an correlation matrix is ​​constructed and then concatenated with the multi-source dataset to obtain a complete input feature set; wherein, the geometric correlation features include beam pointing angle, the gain correlation features include beam gain and main lobe width, the sidelobe correlation features include beam sidelobe power, and the array element phase correlation features include array element real-time phase;

[0010] A two-stage attention-based physically constrained beam prediction model is constructed and trained and validated using the complete input feature set and the corresponding actual beam parameters. The two-stage attention-based physically constrained beam prediction model includes: a feature encoding layer, a two-stage attention layer, a physical constraint decoding layer, and an output layer.

[0011] The predicted beam parameters output by the dual-stage attention-based physical constraint beam prediction model are converted into control signals to drive antenna element adjustment. Based on the link signal-to-noise ratio, bit error rate, and throughput collected by the user terminal, the model parameters are corrected and the beam is split / merged.

[0012] The technical solution provided in this application may include the following beneficial effects:

[0013] This application presents an adaptive beam prediction and control method for satellite mobile communication antennas. This method collects multi-dimensional dynamic data on user motion, signal status, satellite-to-ground attitude, and environmental interference. After time synchronization and preprocessing, it constructs four types of associated features and an association matrix, achieving precise mapping between beam parameters and dynamic scenarios. Furthermore, by leveraging feature encoding, hierarchical attention weighting, and physical rule decoding of a two-stage attention-based physical constraint model, it improves the prediction accuracy and environmental adaptability of parameters such as beam pointing angle and gain. Simultaneously, by converting predicted parameters into control signals to drive real-time adjustment of antenna array elements, and combining link signal-to-noise ratio, bit error rate, and throughput feedback, it completes model parameter correction and beam splitting / merging scheduling, effectively suppressing interference, optimizing resource allocation, and ensuring the stability and transmission efficiency of satellite communication in highly dynamic and complex environments.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0016] Figure 1 A flowchart illustrating the adaptive beam prediction and control method for satellite mobile communication antennas in an exemplary embodiment of this disclosure is shown.

[0017] Figure 2 A detailed flowchart of step S100 of the satellite mobile communication antenna adaptive beam prediction and control method in an exemplary embodiment of this disclosure is shown.

[0018] Figure 3 A detailed flowchart of step S200 of the adaptive beam prediction and control method for satellite mobile communication antennas in an exemplary embodiment of this disclosure is shown.

[0019] Figure 4 This diagram illustrates a two-stage attention-based physical constraint beam prediction model in step S300 of the satellite mobile communication antenna adaptive beam prediction and control method in an exemplary embodiment of this disclosure.

[0020] Figure 5 A detailed flowchart of step S400 of the satellite mobile communication antenna adaptive beam prediction and control method in an exemplary embodiment of this disclosure is shown.

[0021] Figure 6This diagram illustrates a comparison of beam pointing angle errors in different scenarios for the adaptive beam prediction and control method for satellite mobile communication antennas in an exemplary embodiment of this disclosure.

[0022] Figure 7 This diagram illustrates the link quality classification confusion matrix of the adaptive beam prediction and control method for satellite mobile communication antennas in an exemplary embodiment of this disclosure.

[0023] Figure 8 A performance comparison diagram of different monitoring methods for the adaptive beam prediction and control method for satellite mobile communication antennas in the exemplary embodiments of this disclosure is shown. Detailed Implementation

[0024] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0025] This example implementation first provides an adaptive beam prediction and control method for satellite mobile communication antennas. This method can be applied to a terminal device, such as a mobile terminal like a mobile phone, desktop computer, personal digital assistant, laptop, tablet, or smartwatch. (Reference) Figure 1 As shown, the method may include the following steps:

[0026] Step S100: Collect motion data from the user terminal, signal status data from the satellite antenna receiver, satellite attitude data, and environmental interference data from the ground, and perform time synchronization and preprocessing to obtain a multi-source dataset.

[0027] Step S200: Based on the construction of the geometric correlation features, gain correlation features, sidelobe correlation features and array element phase correlation features of the multi-source dataset and beam parameters, an correlation matrix is ​​constructed; wherein, the geometric correlation features include the beam pointing angle, the gain correlation features include the beam gain and the main lobe width, the sidelobe correlation features include the beam sidelobe power, and the array element phase correlation features include the real-time phase of the array elements.

[0028] Step S300: Construct a two-stage attention-based physical constraint beam prediction model, and train and validate it using the multi-source dataset, the correlation matrix, and the corresponding actual beam parameters; the two-stage attention-based physical constraint beam prediction model includes: a feature encoding layer, a two-stage attention layer, a physical constraint decoding layer, and an output layer.

[0029] Step S400: The predicted beam parameters output by the dual-stage attention physical constraint beam prediction model are converted into control signals to drive antenna element adjustment, and the model parameters are corrected and the beam is split / merged based on the link signal-to-noise ratio, bit error rate and throughput collected by the user terminal.

[0030] The aforementioned method can construct four types of correlation features and correlation matrices covering geometry, gain, sidelobes, and array element phase through accurate acquisition and time-synchronized preprocessing of multi-source dynamic data, providing a complete input with clear physical meaning for beam prediction. Then, relying on the two-stage attention physical constraint model, the method processes heterogeneous data through a feature encoding layer, captures spatiotemporal correlations through a two-stage attention layer, and avoids parameter deviations from engineering thresholds through a physical constraint decoding layer, significantly improving the prediction accuracy of parameters such as beam pointing angle and gain. Finally, the predicted parameters are converted into antenna array element control signals, and combined with closed-loop feedback of link signal-to-noise ratio, bit error rate, and throughput, the model parameters and beam splitting / merging strategies are dynamically corrected to effectively suppress the effects of rain attenuation and electromagnetic interference, ensuring the stability and resource utilization of communication in high-dynamic scenarios of low-orbit satellites.

[0031] Below, we will refer to Figures 2 to 8 The steps of the method described above in this example embodiment will be explained in more detail.

[0032] In step S100, motion data from the user terminal, signal status data from the satellite antenna receiver, satellite attitude data, and environmental interference data from the ground are collected, and time synchronization and preprocessing are performed to obtain a multi-source dataset.

[0033] It should be noted that the acquisition equipment and parameters for various data types must be matched to the dynamic characteristics of the corresponding data: user terminal motion data uses a BeiDou BDS-3 / GPS dual-mode positioning module, and its sampling frequency needs to be increased to 5Hz in maritime scenarios; signal status data from the satellite antenna receiver is acquired using a Keysight N9918A spectrum analyzer, and rain attenuation coefficients need to be obtained synchronously for the Ka band; satellite attitude data needs to be acquired in conjunction with an ADIS16488 inertial measurement unit and an array element phase sensor; ground environmental interference data is adapted to a Campbell ARG200 tipping bucket rain gauge and an R&S ESRP3 electromagnetic interference meter. Simultaneously, time synchronization requires a unified clock using an STM32 H743ZI microcontroller to ensure a synchronization error ≤1μs; in the preprocessing stage, outliers are removed using the 3σ criterion and normalization is performed, and the processed multi-source dataset will be stored in the Influx DB time series database.

[0034] In one embodiment, such as Figure 2 As shown, step S100 may include the following sub-steps.

[0035] In step S110, a BeiDou positioning module is integrated into the user terminal to collect the user's three-dimensional coordinates, movement speed, acceleration, and terminal heading angle.

[0036] It should be noted that the user terminal integrates a BeiDou BDS-3 / GPS dual-mode positioning module, which collects the user's three-dimensional coordinates at a sampling frequency of 1Hz. Movement speed and acceleration For maritime users, additional data is collected on the terminal's heading angle. The sampling frequency has been increased to 5Hz, and the data is transmitted back in real time via satellite downlink.

[0037] In step S120, a spectrum analyzer is deployed at the satellite antenna receiver to collect channel state information, signal-to-noise ratio, bit error rate, and carrier frequency offset.

[0038] It should be noted that a Keysight N9918A spectrum analyzer is deployed at the satellite antenna receiver to collect channel state information at a sampling frequency of 10Hz. Signal-to-noise ratio Bit error rate and carrier frequency offset For the Ka band, additional rain attenuation coefficients are collected, and the sampling frequency is synchronized with the user's motion data.

[0039] In step S130, an inertial measurement unit is deployed on the satellite platform to collect the satellite's pitch angle, roll angle, and yaw angle; and a phase sensor is deployed on the antenna array to collect the real-time phase of each array element.

[0040] It should be noted that an ADI ADIS16488 inertial measurement unit is deployed on the satellite platform to collect satellite pitch angles at a sampling frequency of 50Hz. Roll angle Yaw angle Phase sensors are deployed on the antenna array to collect the real-time phase data of each element. , , The number of array elements is determined by the data transmitted to the processing unit via the onboard CAN bus.

[0041] In step S140, a tipping bucket rain gauge is deployed at a ground reference station to collect rainfall and calculate the rain attenuation coefficient, and an electromagnetic interference device is deployed to collect interference signal power and multipath effect intensity.

[0042] It should be noted that a Campbell ARG200 tipping bucket rain gauge was deployed at a ground reference station to collect rainfall, and the rain attenuation coefficient was calculated in conjunction with the interference signal power collected by the R&S ESRP3 electromagnetic interference meter. The sampling frequency is 1Hz; for densely populated urban areas, the multipath effect intensity is additionally collected. It transmits data back to the space-ground collaborative cloud platform via 5G signal.

[0043] In step S150, all device clocks are synchronized, and outliers are removed using the 3σ criterion, and the data is normalized to the [0,1] interval to obtain a multi-source dataset.

[0044] It should be noted that the STM32H743ZI microcontroller is used to uniformly synchronize the clocks of all devices, with a time synchronization error of ≤1μs. The data is preprocessed by the on-board edge computing module (edge ​​gateway). The preprocessing process includes removing outliers using the 3σ criterion and normalizing the data to the [0,1] interval. Then, it is uploaded to the city satellite communication cloud platform through the Ka band and the high-speed satellite-to-ground link at a rate of 1.2Gbps, and finally stored in the time series database InfluxDB.

[0045] In step S200, a correlation matrix is ​​constructed based on the geometric correlation features, gain correlation features, sidelobe correlation features, and array element phase correlation features of the multi-source dataset and beam parameters; wherein, the geometric correlation features include the beam pointing angle, the gain correlation features include the beam gain and the main lobe width, the sidelobe correlation features include the beam sidelobe power, and the array element phase correlation features include the real-time phase of the array elements.

[0046] It should be noted that the beam pointing angle in the geometric correlation feature is not a single parameter, but is adjusted in real time by combining the relative position of the satellite and the ground, the user's motion trend, and the satellite attitude disturbance to adapt to the high dynamic topology characteristics of low-Earth orbit satellites. The beam gain and main lobe width in the gain correlation feature need to be dynamically adjusted in conjunction with the channel signal-to-noise ratio, rain attenuation coefficient, and multipath effect intensity. For example, in urban scenarios with strong multipath effects, the main lobe width will be narrowed accordingly to reduce interference. The sidelobe power in the sidelobe correlation feature needs to establish a constraint relationship with the environmental interference power to avoid sidelobe signals leaking to the direction of the interference source. The real-time phase of the array elements in the array element phase correlation feature needs to establish a precise mapping with the satellite yaw angle to ensure the phase consistency of beamforming. At the same time, the correlation matrix needs to align the feature data of different sampling frequencies with the time step as the dimension. After Min-Max normalization processing, it also needs to pass numerical validity verification and physical logic verification. Finally, it is concatenated with the preprocessed multi-source dataset to form the input feature set for model training in step S300.

[0047] In one embodiment, such as Figure 3 As shown, step S200 may include the following sub-steps.

[0048] In step S210, the basic value of the beam pointing angle is calculated based on the relative position of the satellite and the ground, and the beam pointing angle is corrected based on the motion data of the user terminal and the attitude data of the satellite and the ground.

[0049] It should be noted that this step is the determination step of geometric correlation features. The basic value of beam pointing angle is calculated based on the relative position of the satellite and the ground. At the same time, the user acceleration and heading angle collected by S100 are incorporated to predict the motion trend. The attitude disturbance is compensated by the satellite pitch angle and roll angle. Finally, the corrected pointing angle is output.

[0050] The formula for calculating the basic pointing angle is as follows:

[0051]

[0052] in, The original pitch angle of the beam. The original azimuth angle of the beam. For satellite real-time coordinates, For user coordinates, This is the distance between Earth and space. .

[0053] The formula for correcting the beam pointing angle based on motion data from the user terminal is as follows:

[0054]

[0055]

[0056] in, To correct the pitch angle for the motion trend of the beam user, To correct the azimuth angle for the movement trend of beam users, For acceleration weights, = 0.1, Accelerate for users For the heading angle weight, = 0.05, To predict the step size, = 1, The heading angle at the current moment. The heading angle at the previous moment.

[0057] It should be noted that the acceleration weights = 0.1, Heading Angle Weight = 0.05 is the empirical coefficient based on actual measurements.

[0058] The formula for correcting the beam pointing angle based on satellite-to-ground attitude data is as follows:

[0059]

[0060]

[0061] in, To finally correct the pitch angle of the beam, To finally correct the azimuth angle of the beam, The satellite's elevation angle. This is the satellite's roll angle.

[0062] It should be noted that the coefficient of 0.8 is determined based on the satellite attitude control accuracy, ensuring that the corrected pointing angle error is ≤0.05°.

[0063] In step S220, a mapping relationship between beam gain and channel signal-to-noise ratio is established, and the rain attenuation coefficient is dynamically calibrated by rainfall, frequency loss is compensated by carrier frequency offset, and the main lobe width is narrowed by multipath effect intensity.

[0064] It should be noted that this step is to determine the gain correlation characteristics, based on the ITU-R P.838 rain attenuation model, to establish the relationship between the rainfall collected by S100 and... The quantitative mapping relationship is established, and calibration is performed using real-time rainfall changes. :

[0065] No rain scenario: Rainfall is 0 mm / h. The value is 0.2dB / km, which corresponds to background attenuation on clear days. At this time, the atmospheric attenuation of Ka-band signals is only the basic environmental loss.

[0066] Light rain scenario: Rainfall rate is 0.1-5 mm / h. It changes linearly with increasing rainfall, ranging from 0.8 to 2.5 dB / km, with the value being 0.1 mm / h. =0.8dB / km, at a rainfall of 5mm / h, =2.5dB / km, corresponding to mild rain attenuation, with little impact on link SNR;

[0067] Moderate rain scenario: Rainfall is between 5-15 mm / h. As the rate of increase in rainfall accelerates, the value ranges from 2.5 to 5.0 dB / km, with a rainfall rate of 5 mm / h. =2.5dB / km, with a rainfall of 15mm / h, =5.0dB / km, corresponding to moderate rain attenuation, and the beam gain needs to be adjusted appropriately to compensate for the attenuation;

[0068] Heavy rain scenario: Rainfall rate is 15-30 mm / h. The gradient changes steeply with increasing rainfall, ranging from 5.0 to 8.0 dB / km, with the highest value observed at 15 mm / h. =5.0dB / km, rainfall 30mm / h =8.0dB / km, corresponding to severe rain attenuation, requiring a significant increase in beam gain, while also being wary of the link SNR dropping below 15dB;

[0069] Heavy rain scenario: Rainfall > 30mm / h The attenuation level remains high with increasing rainfall, ranging from 8.0 to 12.0 dB / km, with the highest attenuation at 30 mm / h. =8.0dB / km, at a rainfall of 50mm / h, =12.0dB / km, corresponding to extreme rain attenuation. Gain adjustment alone is insufficient to guarantee link quality, and frequency band switching needs to be triggered.

[0070] The calibration formula for the rain attenuation coefficient is as follows:

[0071]

[0072] in, The calibrated rain attenuation coefficient. This is the initial rain attenuation coefficient. The influence coefficient of rainfall. This represents the rainfall at the current moment. This represents the rainfall at the previous moment.

[0073] The formula for calculating the beam gain is:

[0074]

[0075] in, Based on the beam gain, For signal-to-noise ratio, For satellite launch power, For receiver sensitivity, For free space loss, λ is the wavelength of the Ka band.

[0076] The correction formula for the beam gain is:

[0077]

[0078] in, To correct beam gain, This refers to carrier frequency offset.

[0079] It should be noted that, The unit is kHz. At frequencies above 10kHz, additional compensation gain ensures an SNR ≥ 18dB.

[0080] The correction formula for the multipath interference is:

[0081]

[0082] in, This represents the intensity of the multipath effect.

[0083] It should be noted that when When the value is 5dB, the main lobe width is reduced from 5° to 3°, and the multipath interference attenuation is ≥10dB.

[0084] In step S230, the constraint relationship between beam sidelobe power and interference signal power is defined, and the frequency matching constraint is strengthened by the interference frequency.

[0085] It should be noted that the constraint relationship between the beam sidelobe power and the interference signal power is as follows: the power of the beam sidelobe in the direction of the interference signal incident must be ≤ 1 / 5 of the interference signal power, so as to avoid the sidelobe signal and the interference signal from forming superposition interference; the operation of strengthening the frequency matching constraint by interference frequency refers to the real-time detection of the center frequency of the interference signal. When the frequency deviation between the interference signal frequency and the beam operating frequency is ≤ 5MHz, the carrier frequency offset of the beam is dynamically adjusted to expand the frequency deviation to ≥ 20MHz. At the same time, the bandwidth of the receiving filter corresponding to the beam is narrowed to 1 / 3 of the original bandwidth, thereby improving the isolation of the beam from the interference signal in the frequency dimension and ensuring the effective reception of the main lobe signal.

[0086] The expression for the sidelobe suppression constraint is as follows:

[0087]

[0088]

[0089] in, For sidelobe suppression constraints, For interference frequency overlap, Main lobe power, , The maximum frequency of the interference signal. The minimum frequency of the interference signal. This is the operating frequency band for the beam.

[0090] In step S240, the attitude and array element control are mapped by the satellite yaw angle and the real-time phase of the array element.

[0091] It should be noted that this step is to determine the phase correlation characteristics of the array elements, establish the mapping relationship between attitude and array element control, and ensure beamforming accuracy, that is, to ensure that the phase error is ≤5°.

[0092] The expression for the mapping relationship between attitude and array element control is as follows:

[0093]

[0094] in, The phase of the array element after attitude compensation. For the real-time phase of the array element, Let be the distance between the i-th array element and the center of the antenna. This is the satellite's yaw angle.

[0095] In step S250, the matrix dimension is determined, and the geometric correlation features, gain correlation features, sidelobe correlation features and array element phase correlation features are transformed into a correlation matrix, which is then concatenated with the multi-source dataset to obtain a complete input feature set.

[0096] It should be noted that the matrix dimension is first determined to be M×K, where M is the time step and K is the number of associated features, and the key associated features are identified; the features are then subjected to Min-Max normalization; features with different sampling frequencies are uniformly aligned to a 1-minute time step and padded; an association matrix of 1440×12 is filled according to "time step × feature", and after numerical validity verification and physical logic verification, it is concatenated with the preprocessed original multi-source data to form the model input feature set.

[0097] Specifically, the discrete correlation features of S210-S240 are transformed into structured, time-series model inputs to avoid data dimensional misalignment or disconnection from physical meaning. The specific construction process can be as follows:

[0098] 1. Define the dimension of the correlation matrix as M×K, where M is the time step, and the value of M must match the dynamic monitoring requirements of satellite communication; K is the number of correlation features, and the value of K must cover all key correlation features in S210-S240, ensuring no redundancy or missing features. The specific feature list is as follows:

[0099] Geometric features: Beam pitch angle (°), Beam azimuth angle (°), Earth-Star Distance (km);

[0100] Channel-gain characteristics: beam gain (dB) Free space loss (dB) Rain attenuation coefficient after calibration (dB / km);

[0101] Interference-sidelobe characteristics: sidelobe power (dBm), main lobe power (dBm), Interference signal power (dBm), interference frequency (Hz);

[0102] Attitude-phase characteristics: Corrected element phase (rad), satellite yaw angle (°).

[0103] 2. Due to the significant differences in physical magnitude among different related features, direct concatenation would lead to an over-reliance on a large number of features. Therefore, all features need to be Min-Max normalized to map the values ​​to the [0,1] interval. The normalization formula is as follows:

[0104]

[0105] in, x For the original value of the feature, , These are the minimum and maximum values ​​of the feature in the historical 180-day monitoring data, respectively. After standardization, all features have an equal magnitude of model contribution, avoiding training bias caused by differences in units.

[0106] 3. In step S100, the sampling frequencies of different data sources differ, so all features need to be aligned to a 1-minute time step:

[0107]

[0108] in, This is the completion value at the t-th time step. , These are the measured values ​​at time steps t-1 and t+1, respectively. After alignment, the continuity of the time series needs to be verified to ensure that there are no more than three consecutive missing values ​​in the 1440 time steps; otherwise, the data for that time period should be re-collected.

[0109] 4. Fill the standardized and aligned feature values ​​into a 1440×12 matrix in the order of "time step × feature" to form the correlation matrix. Each element x in the matrix ij This represents the j-th associated feature value (j=1,2,...,12) at the i-th time step (i=1,2,...,1440). After filling, two validations are performed:

[0110] Numerical validity check: Check whether each element is within the normalized interval [0,1]. If it is outside the interval, mark it as an outlier and replace it with the mean of the feature over the next 5 time steps.

[0111] Physical logic verification: Check the physical correlation between features. If it is violated, calculate the deviation value. During subsequent model training, the time step data corresponding to the serious deviation is marked as "key sample" to improve the prediction accuracy in this scenario.

[0112] Final output The matrix must simultaneously meet the requirements of numerical integrity and physical compliance. This matrix will be combined with the original multi-source data after S100 preprocessing to form a complete input feature set for the S300 prediction model, providing a structured and physically meaningful data foundation for the accurate prediction of beam parameters.

[0113] In step S300, a two-stage attention-based physical constraint beam prediction model is constructed and trained and validated using the complete input feature set and the corresponding actual beam parameters. The two-stage attention-based physical constraint beam prediction model includes: a feature encoding layer, a two-stage attention layer, a physical constraint decoding layer, and an output layer.

[0114] Furthermore, such as Figure 4 As shown, the feature encoding layer processes heterogeneous data through a dedicated branch and fuses them into a comprehensive feature tensor; the two-stage attention layer first calculates the feature importance weights of each branch, and then captures long-term temporal dependencies and spatial correlations through a Transformer encoder; the physical constraint decoding layer introduces beam pointing angle constraints, sidelobe suppression constraints, main lobe width constraints, array element phase constraints, and beam gain constraints to correct the decoding process; the output layer outputs the predicted beam parameters through a fully connected layer.

[0115] It should be noted that, firstly, the feature coding layer: a dedicated coding branch is designed for the characteristics of multi-source data to transform heterogeneous data into unified high-dimensional features. User motion feature branch: Uses a 2-layer CNN to process the user coordinates / velocity / acceleration / heading angle time series data from step S100, extracting spatial motion trend features, with an output dimension of 256; Channel state feature branch: Uses a 3-layer GRU to process the SNR / BER / frequency offset / rain attenuation coefficient time series from step S100, capturing the dynamic law of channel fading, with an output dimension of 256; Satellite-ground attitude feature branch: Uses a 2-layer MLP to process the satellite pitch / roll / yaw angle data from step S100, extracting attitude disturbance features, with an output dimension of 128; Environmental interference feature branch: Uses a 2-layer CNN to process the interference signal power / interference frequency / multipath effect intensity / rainfall data from step S100, extracting environmental interference features, with an output dimension of 128; Feature fusion sub-branch: Concatenates the output features from the above branches, with a total dimension of 256+256+128+128=768, and adjusts it to a 512-dimensional comprehensive feature tensor through a 1-layer fully connected layer. .

[0116] Second, the two-stage attention layer: First stage (feature attention): Calculates the importance weights of features in each branch, using the following formula:

[0117]

[0118] in, Assign importance weights to the features of each branch. For the i-th branch feature, the first spatiotemporal weighted feature is output after weighting. .

[0119] Second stage (feature attention): Capturing features based on the Transformer encoder The long-term temporal dependencies and spatial correlations are used to output the second spatiotemporal weighted features. .

[0120] Third, the physical constraint decoding layer: introduces electromagnetic physical constraints to correct the decoding process, avoiding prediction results that violate engineering principles. 0° ≤ ≤60° -180° must be ≤ ≤180°, if outside the range, a penalty of +10 is applied. Predicted values ​​and The deviation must be ≤0.1°; if the deviation exceeds 0.1°, a penalty loss value of +5 is applied; sidelobe suppression constraint: based on Calculate the constraint loss; if violated, the penalty loss value is increased by 5. Main lobe width constraint: the main lobe width must match 2° ≤ main lobe width ≤ 8°; if the deviation exceeds 0.5°, a penalty loss value of +3 is applied. Element phase constraint: the decoded output... It must satisfy -π≤ ≤π, if outside the range, the penalty loss value is increased by 3; Beam gain constraint: decoded output Must meet 25dB≤ ≤40dB; if below 25dB, the penalty loss value is increased by 8. and Auxiliary verification: A three-layer transposed convolution is used. The output of each transposed convolution layer must undergo the above-mentioned constraint verification to ensure that the intermediate results conform to the physical rules of the features. Decoded into a beam parameter feature map.

[0121] Fourth, output layer: The final beam parameters are output through a fully connected layer: the corrected beam pointing angle ( , Corrected array element phase Corrected beam gain The time resolution is 1 minute, the spatial accuracy is 0.1°, and the phase accuracy is 5°.

[0122] In one embodiment, the training strategy for the two-stage attention-based physically constrained beam prediction model is as follows:

[0123] 1. Dataset Construction: 180 days of operational data from BeiDou's low-Earth orbit satellites were collected to construct 4000 sample sets. This includes 3200 training sets and 800 test sets. Each sample set contains preprocessed multi-source data (S100) and an association matrix (S200). (Derived from four types of associated features) and the corresponding beam parameters collected by the satellite telemetry and control system.

[0124] 2. Loss Function Design: A composite loss function is adopted. The formula, which integrates prediction error, physical constraints, and link quality, is as follows:

[0125]

[0126] in, The weights are the root mean square error (RMSE) between the predicted parameters and the true values. =0.4; For physical constraint penalty terms, the weights are... =0.3, which is the sum of the deviation values ​​of each constraint; For SNR-based link quality loss, weights =0.3.

[0127] 3. Phased training:

[0128] Phase 1: Freeze the attention layer and decoding layer, train only the feature encoding layer, and optimize. Convergence threshold <0.05;

[0129] Phase 2: Freeze the encoding layer, train the two-stage attention layer, and optimize. + Convergence threshold <0.03;

[0130] Phase 3: Training the physical constraint decoding layer and output layer, and optimizing them. Convergence threshold <0.02.

[0131] 4. Model Validation: The test set validation results show that the beam pointing angle prediction error is ≤0.1°, the array element phase error is ≤5°, the beam gain error is ≤1dB, and the prediction delay is ≤5ms. Its prediction accuracy and efficiency meet the engineering requirements of dynamic beam control for low-orbit satellites and can be directly used for beam execution control and resource allocation optimization of S400.

[0132] In step S400, the predicted beam parameters output by the dual-stage attention physical constraint beam prediction model are converted into control signals to drive antenna element adjustment, and the model parameters are corrected and the beams are split / merged based on the link signal-to-noise ratio, bit error rate and throughput collected by the user terminal.

[0133] It should be noted that the conversion of the predicted beam parameters into control signals is achieved using a Xilinx Zynq UltraScale+FPGA as the core control platform, along with an AD9739D / A converter and a Hittite HMC649 phase shifter. The control delay must be ≤3ms to match the high dynamic characteristics of low-Earth orbit satellites. The feedback triggering logic for link data needs to be layered according to thresholds: when the link signal-to-noise ratio is <8dB, incremental retraining of the model parameters is directly initiated; when the bit error rate is >10... -4 When the throughput fluctuates by more than ±10%, the beam splitting operation is automatically executed. When the throughput fluctuation is greater than ±10%, the beam combining strategy is adjusted based on the user distribution density (split when the user density is ≥50 users / km², and combine when it is ≤10 users / km²). Simultaneously, beam splitting will be prioritized in emergency communication scenarios to ensure simultaneous access for multiple users. Model parameter corrections must be performed hourly, with incremental updates based on real-time feedback link data to ensure the adaptability and accuracy of beam prediction.

[0134] In one embodiment, such as Figure 5 As shown, step S400 may include the following sub-steps:

[0135] In step S410, the predicted beam parameters output by the dual-stage attention physical constraint beam prediction model are received, and the predicted beam parameters are converted into control signals for the phase shifter / attenuator through internal logic to drive the antenna array elements to adjust.

[0136] It should be noted that the hardware platform of the beam control execution module uses a Xilinx Zynq UltraScale+FPGA as the control core, with a clock frequency of 1GHz, supporting 16-channel array element phase synchronization adjustment; it is equipped with an ADI AD9739 16-bit D / A converter, a Hittite HMC649 phase shifter, and a Mini-Circuits ZX76-250-S+ attenuator. The control logic is that the FPGA receives the beam prediction parameters output from step S300, and converts the parameters into phase shifter / attenuator control signals through internal logic to drive the antenna array element adjustment; the control delay is ≤3ms to ensure matching with the relative motion between the low-Earth orbit satellite and the ground.

[0137] In step S420, the ground user terminal collects and transmits the link signal-to-noise ratio, bit error rate, and throughput in real time, and makes feedback decisions.

[0138] It should be noted that the ground user terminal collects the link signal-to-noise ratio, bit error rate, and throughput in real time and then transmits them back to the on-board processing unit via the satellite downlink.

[0139] Furthermore, the feedback decision is as follows:

[0140] When three consecutive time steps satisfy SNR < 15dB or BER > 10 -4 This triggers the retraining of the prediction model, supplementing the training set with the multi-source dataset mentioned in step S100 in the current scenario and the back-transmitted link signal-to-noise ratio, bit error rate, and throughput, and iteratively training and correcting the parameters of the two-stage attention physical constraint beam prediction model.

[0141] If SNR≥20dB and user density<5 people / km², merge two adjacent beams into one to reduce onboard power consumption;

[0142] If the throughput is less than 10 Mbps and the user density is greater than or equal to 20 people / km², one beam will be split into two smaller beams to increase access capacity.

[0143] In step S430, beam resources are allocated on demand based on user density, interference suppression ratio and the predicted beam parameters.

[0144] It should be noted that the resource allocation in this step needs to be dynamically adapted according to the hierarchical thresholds of user density and interference suppression ratio: For high-density scenarios with a user density ≥ 50 users / km², the existing beam is split into multiple narrow beams based on the high gain parameters of the predicted beam, corresponding to small-scale user clusters to ensure multi-user access; for medium-density scenarios with a user density of 10-50 users / km², the coverage range is finely adjusted based on the main lobe width of the predicted beam to match the current user distribution; for low-density scenarios with a user density ≤ 10 users / km², adjacent beams are merged into wide beams to improve resource utilization. Simultaneously, when the interference suppression ratio is < 15dB, additional anti-interference frequency bands are allocated and the main lobe width is narrowed based on the sidelobe suppression parameters of the predicted beam; when the interference suppression ratio is ≥ 15dB, wideband resources are prioritized to improve the transmission bandwidth of a single beam. The resource allocation scheme is dynamically adjusted every 2 minutes based on real-time updated user density, interference suppression ratio data, and the latest predicted beam parameters to ensure accurate matching of resources with real-time service requirements.

[0145] Furthermore, in high user density scenarios, beam diversity technology is adopted, with a main lobe width of 3° and a gain of 35dB, supporting access for 500 users per beam and a capacity of ≥200Mbps.

[0146] In low user density scenarios, beam combining technology is used, with a main lobe width of 8° and a gain of 28dB, which increases the coverage area of ​​a single beam by 4 times and reduces power consumption to 15W.

[0147] In emergency communication scenarios, high-gain beams and anti-interference frequency bands should be prioritized to ensure emergency data transmission.

[0148] In strong interference scenarios, frequency hopping and beam narrowing are performed, with a hopping frequency interval of ≥50MHz, the main lobe width is reduced to 2°, and the interference suppression ratio is ≥30dB.

[0149] In step S440, communication quality early warning is performed based on the interference signal power, rain attenuation coefficient, multipath effect intensity, and predicted beam parameters.

[0150] It should be noted that the communication quality early warning in this step adopts a three-level classification mechanism. The triggering conditions and corresponding response measures for each level need to be determined in conjunction with the parameter thresholds and predicted beam parameters: Level 1 warning (mild impact) corresponds to interference signal power of -90~-80dBm, rain attenuation coefficient ≤5dB / km, and multipath effect intensity <10dB. At this time, the beam coverage range is finely adjusted in conjunction with the main lobe width parameter of the predicted beam to maintain stable communication quality; Level 2 warning (moderate impact) corresponds to interference signal power of -80~-70dBm, rain attenuation coefficient 5~10dB / km, and multipath effect intensity 10~20dB. The anti-interference algorithm needs to be activated based on the sidelobe suppression parameter of the predicted beam, and the frequency matching relationship of the beam needs to be optimized simultaneously; Level 3 warning (severe impact) corresponds to interference signal power >-70dBm, rain attenuation coefficient >10dB / km, and multipath effect intensity ≥20dB. The frequency band adaptation parameter of the predicted beam needs to be combined to switch to the backup anti-interference frequency band and trigger the beam re-prediction process. Meanwhile, the warning status is dynamically refreshed every 30 seconds based on real-time updated interference and environmental parameters and the latest predicted beam parameters. Furthermore, each level of warning needs to be linked to the resource allocation strategy in step S430 to achieve synchronous adjustment of beam parameters and resource configuration.

[0151] Optionally, the communication quality warning is set to Level 3:

[0152] Mild warning: =5-10dBm or =1-3dB / km or =2-5dB; Adjust the beam direction to avoid interference sources / rain attenuation areas, push notifications to maintenance personnel via APP, and update the warning status every 30 minutes;

[0153] Moderate alert: =10-15dBm or =3-5dB / km or =5-8dB; The anti-interference algorithm is activated, and a text message is sent to the communication operator and the satellite tracking and control center. The announcement screen displays a communication quality prompt.

[0154] Severe Warning: >15dBm or >5dB / km or >8dB; Switch to backup frequency band, activate emergency command system, coordinate with public security, fire department and municipal departments, transmit link data in real time to assist decision-making, update early warning status every 5 minutes until the risk is eliminated.

[0155] Furthermore, the communication performance of the technical solution in this application is analyzed through simulation experiments:

[0156] The experiment selected a low-Earth orbit satellite constellation with a Ka band of 28 GHz and a Ku band of 15 GHz, an orbital altitude of 550 km, 16 antenna array elements, and a beam scanning range of 0-60°. Four typical scenarios were set up, and 100 sets of samples were collected for each scenario, for a total of 400 sets.

[0157] like Figure 6 The comparison of beam pointing angle errors under different scenarios is shown. The pointing angle error of this method is ≤0.1° in all scenarios, with the smallest error at 0.05° in densely populated multi-user scenarios and the largest error at 0.09° in strong electromagnetic interference scenarios. While interference introduces slight noise into the interference frequency data collected in step S100, this error is still controlled within 0.1° through attitude compensation and motion trend prediction. The LSTM prediction error is 0.32-0.48° because it only inputs user position data from step S100 without fusing attitude and interference data, making it unable to handle satellite attitude disturbances and high-speed user movement. The traditional DBF error is 0.51-0.78° because the channel estimation delay cannot match the dynamics of low-Earth orbit satellites, resulting in pointing lag. The fixed beam error is 0.82-1.15° because it lacks dynamic data input from step S100 and relies entirely on static parameters, making it unable to adapt to relative motion between the satellite and the ground and environmental changes. This result confirms the effectiveness of multi-source data fusion and the two-stage attention-physical constraint model, which avoids the limitations of single data-driven or static parameter-dependent approaches by dynamically sensing scene changes and applying electromagnetic physical constraints.

[0158] Table 1

[0159]

[0160] As shown in Table 1:

[0161] The highest SNR of this method is 18.5-24.2 dB: due to the rainfall calibration in the S100 fusion step. Frequency offset correction The beam gain matches the actual channel loss, and the link loss is 3-6dB lower than that of traditional methods.

[0162] Significant advantages in resource utilization: Traditional DBF is 45.8%-58.2%, while the method of this application is 75.2%-88.6%, because this method is based on dynamic beam allocation by user density in step S100;

[0163] The link interruption probability is as low as 1.2-2.5%: it is close to 2.5% only in strong electromagnetic interference scenarios, while the interruption probability of traditional methods is ≥5%. The reason is that this method adjusts the beam and frequency band in advance through interference data in step S100.

[0164] The optimal interference suppression ratio is 28.5-32.1dB, which is more than 20dB higher than that of a fixed beam. This is because the method incorporates the interference frequency overlap calculation in step S100, which strengthens the sidelobe suppression constraint.

[0165] Based on the SNR and BER collected in step S100 and the link quality predicted in step S300, the experiment categorizes the link status into four levels: Excellent (SNR≥20dB, BER≤10) -5 ), good (15≤SNR<20dB, 10 -5 <BER≤10 -4 ), medium (10≤SNR<15dB, 10 -4 <BER≤10 -3 ), difference (SNR<10dB, BER>10) -3 ), construct a confusion matrix to evaluate classification accuracy.

[0166] The results are as follows Figure 7 The image shows a comparison of the confusion matrices for link quality classification across four methods. The accuracy rate for fixed beams is 73.2%. Due to the lack of dynamic data input and reliance solely on static precoding matrices, misclassifications are concentrated in the "Good → Medium" and "Medium → Poor" levels. There are 24 groups of "Actual Good → Predicted Medium" (the most among all methods) and 17 groups of "Actual Medium → Predicted Poor." This is because fixed beams cannot adapt to increased user density or sudden interference, mistakenly classifying "Good state with normal communication" as "Medium state requiring adjustment." This misclassification can lead to either "over-maintenance" or "under-maintenance," wasting satellite resources.

[0167] The accuracy of traditional DBF is 78.8%, relying on real-time channel estimation (CSI) feedback. Due to the 20-30ms delay, it cannot match the dynamics of low-Earth orbit satellites. The misjudgment characteristics correspond perfectly to the "dynamic lag" defect. There are 19 groups of "actual good → predicted medium" and 13 groups of "actual medium → predicted poor". The reason is that the channel estimation delay causes the beam gain adjustment to be untimely. When the SNR drops temporarily due to rain attenuation / multipath, DBF does not compensate for the gain in time, and mistakenly judges the "temporarily fluctuating good state" as the "continuously fading medium state".

[0168] The LSTM prediction accuracy alone was 84.5%. LSTM only inputs user location data from step S100, without incorporating channel, attitude, and interference data. Misclassifications were concentrated in the "Good → Medium" and "Medium → Poor" categories, and while the number of misclassifications was less than the previous two methods, significant shortcomings remained. There were 12 groups of "Actual Good → Predicted Medium" and 15 groups of "Actual Poor → Predicted Medium." This is because LSTM cannot distinguish the "cause of SNR decrease." When SNR decreases, LSTM alone cannot determine whether it is due to user movement or rain attenuation, leading to the misclassification of "medium state due to rain attenuation" as "good state due to movement." This is 11 more misclassifications than the "Good → Medium" method in this application, indicating that "single data-driven" approaches cannot overcome the "multi-factor coupling" problem in link quality classification.

[0169] From the accuracy data, our application achieves 96.5% > LSTM (84.5%) > traditional DBF (78.8%) > fixed beam (73.2%). It is evident that the performance of the four methods exhibits a clear gradient distribution, with our application's accuracy exceeding the other three methods by ≥14 percentage points. Our application outperforms traditional DBF by 18.3 percentage points and fixed beam by 20.2 percentage points, demonstrating the disruptive improvement in link quality classification accuracy achieved through "multi-source data fusion + physical constraints." Furthermore, our application's confusion matrix correctly classifies ≥95 samples on the diagonal, while other methods typically have 65-90 samples on the diagonal, with the diagonal value decreasing as method performance declines. This indicates that our application demonstrates stronger stability in identifying various link quality levels.

[0170] The experiment monitored the long-term performance of this method for 1000 hours. Based on the rainfall, interference power, and user density data collected in real time in step S100 and the beam parameters predicted in step S300, the health index HI was calculated, with 1 being optimal and 0 being failure.

[0171] like Figure 8 As shown, the HI time series curves of four methods (this application, LSTM only, traditional DBF, and fixed beam) intuitively demonstrate the performance stability under long-term dynamic scenarios, supplemented by two threshold lines: health threshold (HI=0.9) and warning threshold (HI=0.85).

[0172] The method described in this application maintains a stable HI between 0.92 and 0.98 without significant attenuation, making it the only method that consistently stays above the healthy threshold (0.9). During a sudden 720-hour rainstorm, when the rainfall attenuation coefficient Ar = 6 dB / km was collected in step S100, the HI briefly dropped to 0.86, approaching the warning threshold, but recovered to 0.93 within 10 minutes through the following mechanism:

[0173] 1. Real-time access to rainfall data from step S100 to calibrate the rain attenuation coefficient;

[0174] 2. Physical constraints force the decoding layer to correct beam gain;

[0175] 3. Closed-loop feedback triggers local retraining of the model.

[0176] Multi-source data fusion provides full-dimensional dynamic perception, a two-stage attention model captures long-term temporal dependencies, physical constraints prevent parameters from deviating from engineering thresholds, and closed-loop control corrects deviations in real time, forming a virtuous cycle of "perception-prediction-execution-feedback".

[0177] LSTM prediction alone: ​​Initial HI=0.92, linearly decreasing to 0.82 after 1000 hours, and approaching the warning threshold (0.85) in the later stages. Relying solely on user location time-series data, without channel / interference / attitude data, it cannot cope with environmental changes such as rain attenuation and electromagnetic interference; furthermore, without physical constraints, the beam gain is mismatched with the actual channel loss, resulting in a continuous decrease in SNR; in addition, without the introduction of closed-loop feedback, the model parameters are fixed and cannot be iteratively optimized with new scenario data.

[0178] Fixed beam: Initial HI=0.85, rapidly decreasing after 200 hours, falling below the warning threshold after 300 hours, and dropping to 0.68 after 1000 hours. Firstly, it relies entirely on the precoding matrix, without any dynamic data input, making it unable to compensate for satellite attitude disturbances and user movement, resulting in a pointing angle error ≥1°. Furthermore, the fixed beam parameters cannot adapt to scenarios such as rain attenuation and dense multi-user environments. In addition, insufficient sidelobe suppression leads to a collapse in link anti-interference capability when electromagnetic interference intensifies later.

[0179] Traditional DBF: The High Intensity Index (HI) remains stable at around 0.88 for the first 400 hours, then rapidly declines, falling below the warning threshold (0.85) at 600 hours, and dropping to 0.72 at 1000 hours. This dynamic lag leads to a performance collapse in later stages. It relies on real-time channel estimation (CSI) feedback, but the 20-30ms delay cannot match the relative motion of low-Earth orbit satellites at 7.8km / s, resulting in beam pointing lag and a pointing angle error ≥0.5°. Furthermore, without user motion trend prediction, the link frequently interrupts in highly dynamic scenarios; additionally, resource allocation is fixed and cannot be adjusted according to user density.

[0180] The 1000-hour verification results show that the stability, anti-interference ability and adaptive adjustment ability of this application are significantly better than the existing technology in long-term dynamic scenarios, providing core technical support for the engineering application of low-orbit satellite communication systems.

[0181] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for adaptive beam prediction and control of satellite mobile communication antennas, characterized in that, include: The system collects motion data from user terminals, signal status data from satellite antenna receivers, satellite attitude data, and environmental interference data from the ground, and performs time synchronization and preprocessing to obtain a multi-source dataset. Based on the construction of geometric correlation features, gain correlation features, sidelobe correlation features, and array element phase correlation features of the multi-source dataset and beam parameters, an correlation matrix is ​​constructed and then concatenated with the multi-source dataset to obtain a complete input feature set; wherein, the geometric correlation features include beam pointing angle, the gain correlation features include beam gain and main lobe width, the sidelobe correlation features include beam sidelobe power, and the array element phase correlation features include array element real-time phase; A two-stage attention-based physically constrained beam prediction model is constructed and trained and validated using the complete input feature set and corresponding actual beam parameters. The two-stage attention-based physically constrained beam prediction model includes: a feature encoding layer, a two-stage attention layer, a physically constrained decoding layer, and an output layer. The feature encoding layer processes heterogeneous data through a dedicated branch and fuses it into a comprehensive feature tensor. The two-stage attention layer first calculates the feature importance weights of each branch, and then captures long-term temporal dependencies and spatial correlations through a Transformer encoder. The physically constrained decoding layer introduces beam pointing angle constraints, sidelobe suppression constraints, main lobe width constraints, array element phase constraints, and beam gain constraints to correct the decoding process. The output layer outputs the predicted beam parameters through a fully connected layer. The predicted beam parameters output by the dual-stage attention-based physical constraint beam prediction model are converted into control signals to drive antenna element adjustment. Based on the link signal-to-noise ratio, bit error rate, and throughput collected by the user terminal, the model parameters are corrected and the beam is split / merged.

2. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 1, characterized in that, The steps of collecting motion data from user terminals, signal status data from satellite antenna receivers, satellite attitude data, and ground environmental interference data, and performing time synchronization and preprocessing to obtain a multi-source dataset include: The user terminal integrates a BeiDou positioning module to collect the user's three-dimensional coordinates, movement speed, acceleration, and terminal heading angle. A spectrum analyzer is deployed at the satellite antenna receiver to collect channel state information, signal-to-noise ratio, bit error rate, and carrier frequency offset. An inertial measurement unit is deployed on the satellite platform to collect the satellite's pitch, roll, and yaw angles; and a phase sensor is deployed on the antenna array to collect the real-time phase of each array element. A tipping bucket rain gauge was deployed at a ground reference station to collect rainfall data and calculate the rain attenuation coefficient. An electromagnetic interference device was also deployed to collect interference signal power and multipath effect intensity. Synchronize the clocks of all devices, and use the 3σ criterion to remove outliers and normalize the data to the [0,1] interval to obtain a multi-source dataset.

3. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 2, characterized in that, The step of constructing geometric correlation features, gain correlation features, sidelobe correlation features, and array element phase correlation features based on the multi-source dataset and beam parameters, and concatenating them with the multi-source dataset after constructing the correlation matrix to obtain a complete input feature set includes: The basic value of the beam pointing angle is calculated based on the relative position of the satellite and the ground, and the beam pointing angle is corrected based on the motion data of the user terminal and the attitude data of the satellite and the ground. Establish a mapping relationship between beam gain and channel signal-to-noise ratio, dynamically calibrate the rain attenuation coefficient based on rainfall, compensate for frequency loss by carrier frequency offset, and narrow the main lobe width by multipath effect intensity. Define the constraint relationship between beam sidelobe power and interference signal power, and strengthen the frequency matching constraint by interfering frequency; By using the satellite yaw angle and the real-time phase of the array elements, a mapping relationship between attitude and array element control is established; Determine the matrix dimension, transform the geometric correlation features, gain correlation features, sidelobe correlation features and array element phase correlation features into a correlation matrix, and concatenate it with the multi-source dataset to obtain the complete input feature set.

4. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 3, characterized in that, The formula for calculating the basic pointing angle is as follows: in, The original pitch angle of the beam. The original azimuth angle of the beam. For satellite real-time coordinates, For user coordinates, This is the distance between Earth and space. ; The formula for correcting the beam pointing angle based on motion data from the user terminal is as follows: in, To correct the pitch angle for the motion trend of the beam user, To correct the azimuth angle for the movement trend of beam users, For acceleration weights, = 0.1, Accelerate for users For the heading angle weight, =0.05, To predict the step size, = 1, The heading angle at the current moment. The heading angle at the previous moment; The formula for correcting the beam pointing angle based on satellite-to-ground attitude data is as follows: in, To finally correct the pitch angle of the beam, To finally correct the azimuth angle of the beam, The satellite's elevation angle. This is the satellite's roll angle.

5. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 4, characterized in that, The calibration formula for the rain attenuation coefficient is: in, The calibrated rain attenuation coefficient. This is the initial rain attenuation coefficient. The influence coefficient of rainfall. This represents the rainfall at the current moment. This represents the rainfall at the previous moment. The formula for calculating the beam gain is: in, Based on the beam gain, For signal-to-noise ratio, For satellite launch power, For receiver sensitivity, For free space loss, λ is the wavelength of the Ka band; The correction formula for the beam gain is: in, To correct beam gain, This refers to carrier frequency offset. The correction formula for the multipath interference is: in, This represents the intensity of the multipath effect.

6. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 5, characterized in that, The expression for the sidelobe suppression constraint is: in, For sidelobe suppression constraints, For interference frequency overlap, Main lobe power, , The maximum frequency of the interference signal. The minimum frequency of the interference signal. The operating frequency band for the beam; The expression for the mapping relationship between attitude and array element control is: in, The phase of the array element after attitude compensation. For the real-time phase of the array element, Let be the distance between the i-th array element and the center of the antenna. This is the satellite's yaw angle.

7. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 2, characterized in that, The steps of converting the predicted beam parameters output by the two-stage attention-based physical constraint beam prediction model into control signals to drive antenna element adjustment, and correcting model parameters and splitting / merging beams based on the link signal-to-noise ratio, bit error rate, and throughput collected by the user terminal, include: The system receives the predicted beam parameters output by the dual-stage attention physical constraint beam prediction model, and converts the predicted beam parameters into control signals for the phase shifter / attenuator through internal logic to drive the antenna array elements to adjust. Ground user terminals collect and transmit link signal-to-noise ratio, bit error rate, and throughput in real time to make feedback decisions. Based on user density, interference suppression ratio and the predicted beam parameters, beam resources are allocated on demand; Based on the interference signal power, rain attenuation coefficient, multipath effect intensity, and predicted beam parameters, a communication quality early warning is issued.

8. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 7, characterized in that, The feedback decision includes: When three consecutive time steps satisfy SNR < 15dB or BER > 10 -4 The multi-source dataset of the current scene and the back-transmitted link signal-to-noise ratio, bit error rate and throughput are added to the training set, and the parameters of the two-stage attention physical constraint beam prediction model are iteratively trained and corrected. If SNR≥20dB and user density<5 people / km², merge two adjacent beams into one to reduce onboard power consumption; If the throughput is less than 10 Mbps and the user density is greater than or equal to 20 people / km², one beam will be split into two smaller beams to increase access capacity.

9. The adaptive beam prediction and control method for satellite mobile communication antennas according to claim 7, characterized in that, The beam resources are allocated on demand, including: In high user density scenarios, beam diversity technology is adopted, with a main lobe width of 3° and a gain of 35dB, supporting access for 500 users per beam and a capacity of ≥200Mbps; In low user density scenarios, beam combining technology is used, with a main lobe width of 8° and a gain of 28dB, which increases the coverage area of ​​a single beam by 4 times and reduces power consumption to 15W. In emergency communication scenarios, high-gain beams and anti-interference frequency bands should be prioritized to ensure emergency data transmission. In strong interference scenarios, frequency hopping and beam narrowing are performed, with a hopping frequency interval of ≥50MHz, the main lobe width is reduced to 2°, and the interference suppression ratio is ≥30dB.

Images