A method for dynamic beam optimization in satellite communication and a satellite communication terminal

Abstract

A method for dynamic beam optimization in satellite communication and a satellite communication terminal are disclosed. The terminal has a locally pre-installed hybrid architecture deep learning model. The specific steps are as follows: The terminal collects spatiotemporal feature data such as space, time, orbit, interference, and position in real time and stores it in a local historical database; it calculates the instantaneous phase angle of the satellite based on the position features and satellite ephemeris, identifies the scene type, and adapts the interference features to obtain optimized features; it extracts historical multi-frame spatiotemporal position features that match the current phase angle, combines them with the optimized features to construct a multi-dimensional feature matrix, inputs it into the model, and outputs spatial and temporal feature scores; it calculates dynamic weights by combining real-time interference and scene, balances the spatiotemporal score decision ratio, integrates historical reputation scores to calculate a comprehensive beam score, selects the optimal beam, and stores the decision result back into the local historical database. This invention integrates spatiotemporal multi-dimensional features, scene-based signal processing, and local deep learning inference to achieve beam optimization, improving the reliability of satellite communication and the efficiency of beam decision-making.

Description

Technical Field

[0001] This invention relates to the field of satellite communication technology, and in particular to a method for dynamic optimization of satellite communication beams and a satellite communication terminal. Background Technology

[0002] Geostationary orbit (GEO) satellite communication systems are widely used in emergency communications and ocean navigation due to their wide coverage and stable deployment. However, influenced by various perturbations, the actual trajectory of GEO satellites exhibits a periodic figure-eight oscillation, causing periodic fluctuations in ground signal quality. Traditional beam selection methods are mainly based on instantaneous received signal strength (RSSI), which has fundamental drawbacks such as a single technical dimension, inability to predict orbital period changes, susceptibility to interference and misleading in complex scenarios such as repeaters, and low decision-making efficiency.

[0003] Existing methods have long average evaluation times and a high probability of service interruption, failing to meet the requirements for highly reliable, real-time communication. Although artificial intelligence (AI / ML) technology is increasingly used in wireless communication, existing solutions have not effectively solved the problem of integrating orbital dynamics prediction and real-time channel awareness, and lack the ability for models to evolve autonomously without relying on terrestrial networks, further limiting the adaptability and long-term reliability of the technical solutions. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings of existing satellite communication beam selection, such as reliance on instantaneous signal strength, single decision-making dimension, and susceptibility to interference in complex scenarios. This invention proposes a dynamic beam optimization method and satellite communication terminal for satellite communication. By integrating spatiotemporal multidimensional features with orbital dynamics prediction, scenario-based signal processing, and local deep learning inference, beam optimization is achieved, thereby improving communication reliability and decision-making efficiency.

[0005] The present invention adopts the following technical solution:

[0006] A satellite communication beam dynamic optimization method, including a locally pre-built hybrid architecture deep learning model on the terminal, further includes the following steps:

[0007] The terminal collects spatiotemporal characteristic data of its environment in real time, including spatial characteristics, temporal characteristics, orbital characteristics, interference characteristics and location characteristics, and stores them in a local historical database;

[0008] Based on the current location features, timestamp, and satellite ephemeris, the terminal calculates the instantaneous phase angle of the satellite, identifies the scene type it is in, and executes an appropriate signal processing strategy for the interference features for different scenes to obtain optimized interference features.

[0009] The orbital features, spatial features, temporal features, and positional features of multiple historical frames that match the current instantaneous phase angle of the satellite are extracted from the local historical database. A multi-dimensional feature matrix is ​​then constructed by combining the optimized interference features. The multi-dimensional feature matrix is ​​then input into the deep learning model to obtain spatial feature scores and temporal feature scores.

[0010] Dynamic weights are calculated by combining real-time interference levels with the scene type. The decision-making proportions of spatial feature scores and temporal feature scores are balanced by the dynamic weights. The comprehensive score of each beam is then calculated by integrating historical reputation scores. The beam with the highest comprehensive score is selected as the optimal beam. The local decision-making results are recorded in the local historical database for subsequent calculation of the historical reputation scores of each beam.

[0011] The formula for calculating the instantaneous phase angle of the satellite is as follows:

[0012] ;

[0013] in, This is the current timestamp. This serves as a reference time for the start of the orbital period. To reference the cumulative number of leap seconds from the previous epoch to the present, For the satellite orbital period, For the initial phase shift, The instantaneous phase angle, This refers to the instantaneous orbital angle shift caused by the perturbation.

[0014] The scenario type includes at least an indoor repeater scenario, and the corresponding signal processing strategy includes: performing interference separation processing to separate the noise introduced by the indoor repeater from the natural multipath interference, and calculating the true signal-to-noise ratio and signal purity; updating the interference features with the true signal-to-noise ratio as the core parameter, and using the signal purity to adjust the interference threshold of the function for calculating the dynamic weight.

[0015] The dynamic weight Calculated using the following adaptive Sigmoid function:

[0016] ;

[0017] in, To measure the interference power, To mitigate measurement uncertainty, For scene-related interference thresholds, Let k be the inherent uncertainty of the system, and k be the slope factor of the Sigmoid function.

[0018] When the scenario is identified as an indoor repeater, the interference threshold Based on the signal purity calculated in this scenario The adjustments will be made dynamically, and the specific rules are as follows:

[0019] when When <0.3, set =-90dBm;

[0020] When 0.3≤ When <0.6, set =-85dBm;

[0021] when When ≥0.6, set = -80dBm.

[0022] The formula for calculating the overall score is as follows:

[0023] ;

[0024] in, For dynamic weights, Scoring the spatial characteristics of beam b Scoring the time characteristics of beam b Historical reputation decay factor Score the historical reputation of beam b.

[0025] The hybrid architecture deep learning model includes a spatial feature extraction part and a temporal feature extraction part:

[0026] The spatial feature extraction part adopts a combination structure of convolutional neural network (CNN) and long short-term memory network (BiLSTM) to extract spatial correlation features and output spatial feature scores.

[0027] The time feature extraction part adopts a Transformer network structure to extract time series features and output time feature scores.

[0028] The deep learning model is updated using a satellite channel-based federated learning mechanism, including:

[0029] The terminal performs model fine-tuning based on data from the local historical database, generates gradient update data, performs multi-level compression processing on the gradient update data, and uploads the multi-level compressed gradient update data to the satellite ground station via the satellite channel.

[0030] The satellite ground station securely aggregates gradient update data from multiple terminals to generate a global model update, which is then broadcast and distributed via satellite channel.

[0031] The terminal receives and applies the global model update.

[0032] The multi-level compression includes gradient sparsity processing, non-uniform integer quantization, differential coding, and entropy coding compression; the secure aggregation includes first filtering the gradient data for abnormal gradients to remove invalid and abnormal gradient information, then performing the aggregation operation by adding differential privacy noise and using homomorphic encryption, and simultaneously performing weighted aggregation based on the local data quality and data volume of each terminal.

[0033] A satellite communication terminal, comprising:

[0034] The spatiotemporal feature acquisition module is used to collect spatiotemporal feature data of the surrounding environment in real time, including spatial features, temporal features, orbital features, interference features, and location features.

[0035] The feature construction module calculates the instantaneous phase angle of the satellite based on the current position features, timestamp, and satellite ephemeris, identifies the scene type, and performs adaptive signal processing strategies on the interference features for different scenes to obtain optimized interference features; it extracts the orbit features, spatial features, temporal features, and position features of historical multiple frames that match the current instantaneous phase angle of the satellite from the local historical database, and constructs a multi-dimensional feature matrix by combining the optimized interference features;

[0036] The beam scoring selection module inputs the multi-dimensional feature matrix into the hybrid architecture deep learning model to perform local inference, and outputs spatial feature scores and temporal feature scores respectively; it calculates dynamic weights by combining real-time interference level and scene type, and integrates historical reputation scores to obtain a comprehensive score for each beam; finally, it selects the beam with the highest comprehensive score as the optimal beam, and records the local decision results to the local historical database for subsequent calculation of historical reputation scores for each beam.

[0037] A local historical database is used to store the spatiotemporal feature data and to store a pre-built deep learning model with a hybrid architecture.

[0038] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. In this invention, the terminal has a locally pre-installed hybrid architecture deep learning model. It collects and stores spatial, temporal, orbital, interference, and location spatiotemporal features from all dimensions to form a local basic database. It calculates the instantaneous phase angle of the satellite by combining the position features and satellite ephemeris. At the same time, it completes scene recognition and performs scene-specific adaptation and optimization of interference features. By matching the historical multi-frame features of the current phase angle, a multi-dimensional feature matrix is ​​constructed to ensure that the feature input conforms to the dynamic change law of the satellite orbit. Then, through model inference, dynamic weight balancing, and fusion of historical reputation, the beam comprehensive scoring and optimization are completed. The decision results are stored back in the local historical database to form a data closed loop. The entire process does not rely on the ground network, which improves the terminal's independent operation capability in satellite communication scenarios.

[0040] 2. In this invention, the instantaneous phase angle of the satellite is calculated using a specific formula. The formula incorporates key influencing factors such as orbital period, cumulative leap seconds, initial phase offset, and orbital angle offset caused by perturbation. This effectively offsets various deviations in the satellite's orbital operation, improves the accuracy of phase angle calculation, ensures that the historical features extracted subsequently are highly matched with the current satellite orbital state, avoids feature mismatch problems caused by phase angle deviation, and guarantees the effectiveness of the multi-dimensional feature matrix.

[0041] 3. In this invention, a dedicated signal processing strategy is designed for indoor repeater scenarios. An interference separation algorithm is used to separate noise introduced by the repeater from natural multipath interference. The true signal-to-noise ratio is calculated and integrated into the feature system to complete interference feature optimization. At the same time, the interference threshold of the function for calculating dynamic weights is adaptively adjusted based on the signal purity value range, and the signal purity is incorporated into the beam comprehensive scoring system. This effectively offsets the negative impact of signal distortion and interference on decision-making in repeater scenarios, filling the gap in the adaptability of traditional beam optimization methods in complex scenarios.

[0042] 4. In this invention, the model is divided into spatial and temporal feature extraction branches. The spatial branch adopts a CNN+BiLSTM combination structure to mine the spatial correlation features of the beam signal; the temporal branch adopts a Transformer network structure to efficiently extract the temporal series features of the beam signal. The two branches perform their respective functions and output independent scores, realizing comprehensive and targeted mining of beam features. This solves the problem that a single network structure is not sufficient for mining spatiotemporal features, and provides accurate feature basis for subsequent beam scoring.

[0043] 5. In this invention, the model is updated by relying on the federated learning mechanism of the satellite channel. The terminal locally fine-tunes the gradient data and compresses it through multiple levels to adapt to the satellite channel transmission characteristics. The ground station generates a global update through multi-stage secure aggregation, which not only ensures that the model iteration fits the actual application scenario of the terminal, but also takes into account the transmission efficiency and update security.

[0044] 6. In this invention, dynamic weights are calculated by an adaptive Sigmoid function, and the decision weight ratio of spatiotemporal feature scores is combined with the real-time interference level and scene type. Then, a comprehensive score is calculated by a specific formula and historical reputation scores are integrated. This takes into account both the current state of the beam and its long-term performance, avoids decision bias caused by single data fluctuations or fixed weights, and improves the robustness of beam optimization decisions. Attached Figure Description

[0045] Figure 1 This is the main flowchart of the method of the present invention;

[0046] Figure 2 A schematic diagram of the phase partitioning of the figure-eight trajectory of a geostationary satellite.

[0047] Figure 3 This is a structural diagram of the hybrid architecture deep learning model (CNN-BiLSTM-Transformer) of this invention.

[0048] Figure 4 This is a timing diagram of the update of the federated learning model based on satellite channels in this invention.

[0049] Figure 5 This is a graph showing how the dynamic weight adjustment of the present invention changes with the level of interference.

[0050] Figure 6 This is a schematic diagram of the hardware structure of the satellite communication terminal of the present invention.

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0052] The present invention will be further described below through specific embodiments.

[0053] This invention proposes a dynamic beam optimization method for satellite communication, utilizing a locally pre-built hybrid architecture deep learning model on the terminal to achieve deep fusion inference of spatiotemporal features. The terminal is a satellite communication terminal with fully autonomous decision-making capabilities, comprising an RF and baseband module, a positioning and timing module, a computing and storage module, and a power supply module, etc. See also... Figure 1 The method in this embodiment specifically includes the following steps:

[0054] The S1 terminal collects real-time spatiotemporal characteristic data of its surrounding environment, including spatial characteristics, temporal characteristics, orbital characteristics, interference characteristics, and location characteristics, and stores them in a local historical database.

[0055] This step is the foundational data support for dynamic beam optimization. Its core purpose is to acquire full-dimensional feature data covering space, time, orbit, interference, and location through parallel acquisition by multiple source sensors, providing a high-quality data source for subsequent steps. All acquired data is stored in a local historical database after spatiotemporal alignment to ensure the data independence of the terminal's autonomous decision-making.

[0056] The spatial feature focusing signal instantaneous physical layer characteristics specifically include the received signal strength (RSSI_Kalman) after Kalman filtering, signal-to-noise ratio variance (σ²_SNR), Doppler frequency offset (f_Doppler), and multipath angle spread (σ_θ^AS). By scanning the preset beam center frequency through the terminal's RF and baseband modules, IQ data is collected at each frequency point. Signal power, subband power ratio, and multipath-related parameters are calculated in real time to characterize signal strength stability, motion characteristics, and the impact of multipath interference, providing a fundamental basis for real-time beam quality assessment.

[0057] Temporal features are extracted based on the temporal evolution of the signal, covering the 24-hour periodic variation amplitude (A_24h), the 6-hour periodic variation amplitude (A_6h), and the instantaneous trend of signal-to-noise ratio (dSNR / dt). Based on historical signal data from multiple consecutive acquisition cycles, the periodic fluctuation amplitude and slope are calculated using time-series analysis methods to capture the long-term trend and instantaneous fluctuation characteristics of the signal over time.

[0058] Orbital characteristics are related to satellite orbital dynamics, the core of which include the satellite's instantaneous phase angle (θ(t)), the rate of change of satellite-to-ground distance (dr / dt), and the rate of change of orbital eccentricity (de / dt). See also Figure 2 By reading the locally stored satellite ephemeris and perturbation correction table, and combining the current timestamp and terminal location information, the orbital state parameters of the satellite on the figure-eight trajectory can be calculated. At the same time, four perturbation terms, namely lunar gravity, solar gravity, Earth oblateness and solar radiation pressure, are introduced to correct the satellite's orbital motion state.

[0059] Interference characteristics are used to quantify the current electromagnetic interference level, specifically including the total in-band interference power (I_in-band) and the estimated signal-to-interference-plus-noise ratio (SINR_estimated). A full-band FFT scan of the operating frequency band is performed using the RF front-end's spectrum analysis mode to identify and integrate the interference peak power. The SINR is then calculated by combining the signal power with the system's thermal noise floor, thus determining the interference intensity at the current beam frequency.

[0060] Location features are centered on BeiDou positioning information, including latitude and longitude grid hash, normalized altitude (Alt_norm), satellite relative elevation (Elevation), and historical beam preference index (BeamPref). The positioning and timing module parses BeiDou positioning messages to obtain raw data such as latitude, longitude, altitude, and time. After encoding and normalization, this data is combined with local historical database queries to generate a historical beam preference index, providing a geographic reference for the terminal.

[0061] Based on the current location features, timestamp, and satellite ephemeris, the S2 terminal calculates the instantaneous phase angle of the satellite, identifies the type of scene it is in, and executes an appropriate signal processing strategy for the interference features in different scenes to obtain optimized interference features.

[0062] The instantaneous phase angle is a core parameter for quantifying the instantaneous position of a GEO satellite's figure-eight trajectory. Its calculation is based on satellite ephemeris, terminal position characteristics, and UTC timestamps, while incorporating four perturbation terms to ensure accuracy. The specific calculation formula is as follows:

[0063] ;

[0064] in, This is the current timestamp; This serves as the reference time for the start of the orbital period. This represents the cumulative number of leap seconds from the initial reference time to the current time. For the satellite orbital period, For the initial phase shift, The instantaneous phase angle, The instantaneous orbital angle deviation caused by the perturbation is shown in Table 1.

[0065]

[0066] in The perturbations include lunar gravity, solar gravity, Earth's oblateness, and solar radiation pressure. These perturbations are obtained from a pre-set perturbation correction table on the terminal. The parameters of each perturbation source are shown in Table 2.

[0067]

[0068] In this embodiment, the scene type is determined by jointly analyzing multi-dimensional signal indicators, including open field scenes, densely populated urban areas, and indoor repeater scenes. During the identification process, key indicators such as the matching relationship between received signal strength and signal-to-noise ratio, the distribution of narrowband interference peaks in the spectrum, the degree of multipath delay spread, and signal envelope distortion characteristics are considered. When the characteristic judgment conditions of the corresponding scene are met, the scene type is determined.

[0069] Taking an indoor repeater scenario as an example, its determination requires meeting at least three of the following characteristics: Received Signal Strength Indicator (RSSI) higher than -90dBm but Signal-to-Noise Ratio (SNR) lower than 8dB, indicating an abnormal match; the ratio of signal strength to SNR exceeding three times the typical value in open environments; a significant narrowband interference peak appearing in the spectrum at specific frequency points (e.g., 1.8MHz, 3.6MHz); multipath delay spread greater than 40 nanoseconds; and the signal envelope exhibiting a "flat-top" distortion characteristic. Through real-time detection and joint analysis of the above multi-dimensional signal indicators by the terminal, if the determination conditions are met, it is determined to be an indoor repeater scenario. For open area scenarios and densely populated urban areas, the determination is also based on the spatial characteristics, interference characteristics, and location characteristics collected in step S1, and each scenario only needs to meet the corresponding characteristic conditions.

[0070] For different scenarios such as open areas, densely populated urban areas, and indoor repeaters, the terminal implements differentiated adaptation strategies for interference features to eliminate scenario-specific interference and improve feature reliability. Specifically, based on the differences in the electromagnetic environment of the scenario, targeted interference identification, separation, or suppression algorithms are used to filter out scenario-specific interference components, ultimately obtaining optimized interference features that are highly adapted to the current scenario.

[0071] Taking an indoor repeater scenario as an example, the terminal executes the corresponding signal processing strategy: based on the interference characteristics, it estimates the power spectral density of the interference noise through adaptive filtering or blind source separation technology. Calculate the true signal-to-noise ratio The ideal channel response is estimated using a known spreading pilot sequence. Noise introduced by the indoor repeater and natural multipath interference are then separated from the total channel response, and the signal purity is calculated. The estimated signal-to-interference-plus-noise ratio (SINR_estimated) is updated using the true signal-to-noise ratio as the core parameter.

[0072] True signal-to-noise ratio The calculation formula is as follows:

[0073] ;

[0074] in, The total power of the received signal. The separated repeater interference power. This represents the background thermal noise power of the terminal receiver.

[0075] Signal purity The calculation formula is as follows:

[0076] ;

[0077] This represents the ideal channel response for channel estimation using a known pilot sequence. To estimate the total channel response from the actual received signal, The response for channel estimation of non-repeater multipath.

[0078] S3 extracts orbital, spatial, temporal, and positional features from multiple historical frames that match the current satellite's instantaneous phase angle from the local historical database. It then combines these features with optimized interference features to construct a multi-dimensional feature matrix. The multi-dimensional feature matrix is ​​then input into a deep learning model to obtain spatial and temporal feature scores.

[0079] In this step, matching features are extracted from the local historical database and a multi-dimensional feature matrix is ​​constructed. The core is to use the periodicity of satellite orbit phase to associate historical data with the current state. Specifically, the terminal first uses the currently calculated instantaneous satellite phase angle θ(t) as a reference to query the most recent historical multi-frame data in the local historical database that is close to the current orbit position, and extracts the spatial, temporal, and positional features of the corresponding frames to ensure that the extracted historical features can reflect the signal patterns under the same orbital state. Then, these three types of historical features are integrated with the interference features optimized in step S2 to finally construct a sixteen-dimensional multi-dimensional feature matrix covering five sub-vectors: space (4-dimensional), time (3-dimensional), orbit (3-dimensional), interference (2-dimensional), and position (4-dimensional), i.e., a sixteen-dimensional feature vector F, expressed as follows:

[0080] F=[F_spatial,F_temporal,F_orbit,F_interference,F_location]^T;

[0081] Wherein, F_spatial represents the spatial feature vector, F_temporal represents the temporal feature vector, F_orbit represents the orbital feature vector, F_interference represents the interference feature vector, and F_location represents the location feature vector. The composition of each sub-vector is shown in Table 3.

[0082]

[0083] In this embodiment, the hybrid architecture deep learning model is designed for deep mining of spatiotemporal features of multi-dimensional feature matrices. Based on a lightweight optimization scheme deployed locally on the terminal, it includes spatial feature extraction and temporal feature extraction parts. A dual-branch parallel structure extracts spatial and temporal features respectively and outputs corresponding scores, providing core reasoning basis for subsequent beam optimization and comprehensive decision-making. Figure 3 Specifically as follows:

[0084] The spatial feature extraction section extracts spatial correlation features and outputs spatial feature scores. A combined structure of a Convolutional Neural Network (CNN) and a Long Short-Term Memory (BiLSTM) network is employed to mine the correlation between spatial and positional features in the feature matrix: local spatial features are extracted through the convolutional layers of the CNN network, capturing the spatial mapping relationship between multi-beam received signal strength, multipath angular spread characteristics, and terminal location and signal quality, while filtering out redundant noise in single-frame spatial data; then, the BiLSTM network processes the spatial feature sequences of multi-frame historical matching data, preserving the continuity of spatial features as the frame order changes, compensating for the CNN's deficiency in capturing temporal correlations, and finally outputting a quantified spatial feature score. It represents the overall level of beam space transmission quality under current and historical matching conditions.

[0085] The CNN network includes three progressive one-dimensional convolutional layers, which generate multi-dimensional feature matrix vectors. Using this as input, by progressively expanding the receptive field and increasing the feature dimensionality, we can mine local spatial correlations and generate high-dimensional abstract features, as detailed below:

[0086] The first convolutional layer uses a kernel of length 3 to map the 16-dimensional input to 32-dimensional features and introduces non-linear activation: ;

[0087] The second convolutional layer increases the kernel length to 5, capturing a wider range of spatial relationships with a stride of 5, further expanding the receptive field and increasing the feature dimension to 64: ;

[0088] The third convolutional layer further increases the kernel length to 7, capturing global spatial relationships with a stride of 7, and finally outputs 128-dimensional high-dimensional abstract spatial features, providing a spatial feature foundation for subsequent BiLSTM layers to process temporal relationships: ;

[0089] This is a one-dimensional convolution operation. , , For the output of the convolutional layer, It is a non-linear activation function. , For weight tensors, It is the set of real numbers.

[0090] The Bidirectional Long Short-Term Memory (BiLSTM) network takes 1×128-dimensional high-dimensional spatial features output from three layers of one-dimensional convolutions as input. After passing through forward and backward LSTM units with a hidden dimension of 128 to capture the bidirectional temporal correlation of spatial features, it outputs a 1×256-dimensional feature vector. This vector will be used in subsequent fully connected layers and Sigmoid activation to generate spatial feature scores. .

[0091] The time feature extraction part employs a Transformer network structure to extract time series features and output a time feature score. It focuses on mining optimized orbital features, time features, and historical temporal patterns from the feature matrix. Leveraging the Transformer's multi-head self-attention mechanism, it captures the signal temporal dependencies caused by periodic changes in satellite orbits. Simultaneously, it incorporates the satellite's instantaneous phase angle into position encoding, deeply binding time series feature extraction with satellite orbital state. This effectively identifies the correlation between orbital phase changes and signal quality fluctuations, ultimately outputting a time feature score. ), characterizing the trend level of beam quality evolution with orbital timing.

[0092] Specifically, the following steps are used to achieve deep binding between orbital phase and temporal features, ultimately outputting a quantified temporal feature score:

[0093] First, the sixteen-dimensional feature vector F is added to the position code generated based on the satellite's instantaneous phase angle to obtain the feature vector fused with orbital timing information:

[0094] ;

[0095] Then The input is fed into a Transformer encoder layer containing four self-attention heads, where each head represents the query dimension. AND dimension Set as Attention score is calculated using the following formula:

[0096] ;

[0097] These represent the query, key, and value vectors, respectively. `Attention()` is the attention function. The core of `Attention(Q,K,V)` is to calculate the similarity between the query vector (representing the current feature state) and the key vector (representing historical feature states), generate a set of attention weights (normalized via softmax), and then use these weights to perform a weighted sum on the value vector (representing effective information from historical states). The 4-head parallel structure can capture temporal dependencies in different dimensions. This is used to alleviate the vanishing gradient problem caused by excessively high attention scores; self-attention output and After performing residual connections, intermediate features are obtained through layer normalization:

[0098] ;

[0099] The intermediate feature Z is input into a two-layer fully connected feedforward network FFN (dimensionality transformed from 16→64→16), which enhances the feature representation capability through nonlinear transformation; the output is then residually connected to Z and normalized to obtain the final encoder output.

[0100] ;FFN is a two-layer fully connected network (16→64→16).

[0101] Finally, Z′ is input into a fully connected layer, and after being mapped to the [0,1] interval by a Sigmoid activation function, the temporal feature score is obtained:

[0102] .

[0103] This represents the Sigmoid activation function, which maps any real-valued input x to the interval (0,1). Specifically:

[0104]

[0105] It is a learnable parameter matrix that is responsible for linearly combining (or weighting and summing) 16-dimensional temporal high-level features into a single-dimensional pre-score.

[0106] It is a scalar bias, which is a linear transformation. Provide a reference offset.

[0107] In this embodiment, the deep learning model is updated through a satellite channel-based federated learning mechanism, which provides an online update method for the model. The update of the federated learning model is performed via OTA (Over-The-Air) communication via satellite communication link. The normal update cycle is set to be no less than 30 days. During the interval between two model updates, the terminal can independently complete the relevant business processing by relying entirely on the local pre-built model and historical database, ensuring the normal operation of the system.

[0108] Specifically, the terminal performs lightweight fine-tuning training on the locally pre-built deep learning model based on beam-related historical data in the local historical database to generate gradient update data. To adapt to the transmission characteristics of the satellite channel, the terminal performs multi-level compression processing on the gradient update data, including gradient sparsification, non-uniform integer quantization, differential coding, and entropy coding compression, reducing the amount of data transmission while ensuring the effectiveness of model updates. Subsequently, the multi-level compressed gradient update data is uploaded to the satellite ground station through the satellite communication link.

[0109] The satellite upload step incorporates a channel quality adaptive scheduling strategy: the terminal monitors the current satellite channel quality in real time, selects idle service periods to initiate gradient data upload requests, dynamically adjusts transmission parameters such as modulation and coding scheme and transmission rate according to channel quality, and supports a breakpoint resume mechanism to avoid transmission interruptions and data retransmission losses caused by channel fluctuations.

[0110] After collecting gradient update data uploaded by multiple terminals, the satellite ground station initiates a secure aggregation process to securely aggregate the gradient update data from multiple terminals. First, abnormal gradient filtering is performed on the gradient data to remove invalid and abnormal gradient information. Then, differential privacy noise is added and homomorphic encryption is used to perform the aggregation operation. At the same time, weighted aggregation is performed by combining the local data quality and data volume of each terminal. After multiple security processing steps, a global model update is generated and broadcast to all terminals in the network via satellite channel.

[0111] After receiving the global model update data broadcast by the satellite ground station via the satellite channel, each terminal completes the adaptation, update and iterative optimization of the local deep learning model, thereby achieving continuous improvement in model performance.

[0112] In this embodiment, the online model update mechanism is highly adapted to the characteristics of satellite channel transmission. Based on the federated learning update method of satellite channel, it effectively adapts to the characteristics of satellite channel transmission and reduces data transmission loss through multi-level compression of gradient update data and adaptive scheduling of channel quality in the satellite upload process. At the same time, it relies on security aggregation strategies such as adding differential privacy noise and homomorphic encryption aggregation to ensure the security and effectiveness of model update. The setting of the normal update cycle also takes into account the model iteration efficiency and terminal operating load. Furthermore, the method of locally fine-tuning the generation of gradient data on the terminal makes the model update more in line with the actual application scenario of the terminal.

[0113] S4 calculates dynamic weights by combining real-time interference levels and scene types. It balances the decision-making proportions of spatial feature scores and temporal feature scores through dynamic weights, and then integrates historical reputation scores to calculate the comprehensive score of each beam. The beam with the highest comprehensive score is selected as the optimal beam, and the local decision results are recorded in the local historical database for subsequent calculation of the historical reputation score of each beam.

[0114] In this embodiment, the dynamic weights are calculated using an adaptive Sigmoid function, and the specific calculation formula is as follows:

[0115] ;

[0116] in, For dynamic weights, To measure the interference power, To mitigate measurement uncertainty, For scene-specific interference thresholds (e.g., for repeater scenarios) =-85dBm), The system inherent uncertainty is given by k, which is the slope factor of the Sigmoid function; the parameters are explained in Table 4.

[0117]

[0118] The extreme cases are as follows:

[0119] when ;

[0120] when (Completely uncertain, take the median value);

[0121] The dynamic weight adjustment mechanism establishes the following relationship between interference level and weight:

[0122] ;

[0123] See Figure 5 The physical meaning of the dynamic weight adjustment curve:

[0124] High interference scenarios For applications such as repeater stations and densely populated urban areas, real-time measurements are believed to be more reliable.

[0125] Low interference scenarios For open areas and suburbs, the trajectory prediction model is believed to be more accurate.

[0126] The partial derivative analysis (sensitivity) is as follows:

[0127] ;

[0128] Physical meaning: Measurement uncertainty The larger the value, the higher the weight. Interference The lower the sensitivity to change.

[0129] In this embodiment, in the indoor repeater scenario, the interference threshold in the dynamic weight calculation formula is... Based on the signal purity calculated in this scenario Dynamic adjustments are made, specifically including: setting the value of the interference threshold based on the numerical range of signal purity, so as to balance the decision-making proportion of spatial feature scoring and temporal feature scoring;

[0130] When signal purity When <0.3, set =-90dBm;

[0131] When 0.3 ≤ signal purity When <0.6, set =-85dBm;

[0132] When signal purity When ≥0.6, set = -80dBm.

[0133] In this embodiment, the decision-making proportions of spatial feature scores and temporal feature scores are dynamically weighted and balanced, and then the comprehensive score of each beam is calculated by fusing historical reputation scores. The specific formula is as follows:

[0134] ;

[0135] in, For dynamic weights, Scoring the spatial characteristics of beam b Scoring the time characteristics of beam b Historical reputation decay factor Score the historical reputation of beam b.

[0136] Historical reputation score is an important dimension in the calculation of beam comprehensive score. This score is calculated based on the historical service performance of each beam under the same or similar satellite orbit phase and similar communication scenarios. The core reference beams have key indicators such as communication stability, signal quality compliance rate, interference tolerance, and link availability during their historical operation. The standardized score results are formed through quantitative statistics and normalization. The score directly reflects the reliability and reputation level of the corresponding beam's historical service.

[0137] Specifically, historical credit rating The calculation formula is as follows:

[0138] ;

[0139] in, The time decay weight is calculated using the following formula:

[0140] .

[0141] The specific parameters are explained in Table 5 below:

[0142]

[0143] The method in this application makes beam optimization decisions that fit the actual characteristics of satellite communication scenarios. By extracting historical features that match the instantaneous phase angle of the current satellite, and combining them with scenario-optimized interference features to construct a multi-dimensional feature matrix, the method relies on a deep learning model with a CNN+BiLSTM and Transformer dual-branch hybrid architecture to mine the spatial correlation and temporal series features of the beam signal. Then, by dynamically balancing the spatiotemporal feature score ratio and integrating the beam's historical reputation score, a multi-dimensional comprehensive beam score calculation is achieved. This ensures that beam optimization decisions not only fit the current communication status but also take into account historical performance, thereby improving the rationality and accuracy of the decisions.

[0144] Furthermore, to meet the deployment and operational requirements of satellite terminals, the terminal has a pre-installed deep learning model and historical database. During the interval between two OTA updates of the federated learning model, it can independently complete beam optimization-related tasks without relying on continuous network interaction.

[0145] In this embodiment, a satellite communication terminal is also proposed to implement the aforementioned satellite communication beam dynamic optimization method. This terminal is adapted to satellite communication beam dynamic optimization scenarios and achieves a closed-loop operation of the entire process—feature acquisition, processing, inference decision-making, and data storage—through modular division of labor. The functions and collaborative logic of each module are as follows:

[0146] The spatiotemporal feature acquisition module is used to collect spatiotemporal feature data of the surrounding environment in real time. This includes spatial features (such as multi-beam received signal strength and multipath angular spread characteristics), temporal features (such as data acquisition time stamps and satellite orbital periodicity information), orbital features (real-time satellite orbital parameters obtained based on satellite ephemeris), interference features (such as interference signal type, intensity, and frequency distribution), and location features (terminal's own positioning information). The acquisition process balances real-time performance with data integrity, providing accurate raw data support for subsequent feature processing and model inference.

[0147] The feature construction module calculates the satellite's instantaneous phase angle based on the current location features, timestamp, and satellite ephemeris, identifies the scene type, and applies appropriate signal processing strategies to the interference features for different scenes. For example, in complex interference scenes, it enhances the filtering and noise reduction processing of interference features, and finally outputs optimized interference features to improve feature effectiveness. It extracts historical multi-frame data that matches the current satellite's instantaneous phase angle from the local historical database, extracts the orbital features, spatial features, temporal features, and position features of the corresponding frames, and constructs a multi-dimensional feature matrix by combining the optimized interference features.

[0148] The beam rating selection module inputs a multi-dimensional feature matrix into a hybrid architecture deep learning model to perform local inference, and outputs spatial feature scores and temporal feature scores respectively. It calculates dynamic weights by combining real-time interference levels and scene types, and integrates historical reputation scores to obtain a comprehensive score for each beam. Finally, it selects the beam with the highest comprehensive score as the optimal beam, and records the local decision results to the local historical database for subsequent calculation of historical reputation scores for each beam.

[0149] The local historical database, serving as the core carrier for terminal data storage and model support, stores spatiotemporal feature data and pre-built hybrid architecture deep learning models. On one hand, this database stores raw data continuously collected by the terminal's spatiotemporal feature acquisition module, beam decision records, and beam performance data, gradually accumulating into a structured local historical database. This provides a data source for historical feature extraction by the feature construction module and historical reputation retrieval by the beam scoring selection module. On the other hand, it stores pre-built hybrid architecture deep learning models, supporting local inference operations of the beam scoring selection module and ensuring the terminal can still operate independently even without network update interactions.

[0150] Application Examples

[0151] The terminal in this embodiment also includes the following hardware modules: a radio frequency and baseband module, a positioning and timing module, a computing and storage module, and a structure and power supply module, etc., see [link to documentation]. Figure 6 Taking the Tiantong-1 satellite communication system and a dedicated terminal equipped with Huawei's Ascend 310 AI chip as examples, the specific examples are as follows:

[0152] As the core of terminal signal transceiver and physical layer processing, the radio frequency and baseband modules are precisely configured to meet the requirements of satellite communication signal transmission. The antenna layer adopts an S-band duplex microstrip patch antenna with a gain of 10dBi. The front end of the radio frequency processing layer is equipped with an ADIAD9361 chip, which supports wideband operation from 70MHz to 6GHz with a bandwidth of <56MHz. The baseband processor is a Xilinx Zynq-7030, which completes the entire process of the satellite communication physical layer protocol stack.

[0153] The positioning and timing module provides the terminal with high-precision location information and time reference, ensuring the accuracy of satellite phase angle calculation and scene recognition. The antenna layer also adopts an L-band Beidou dual-frequency receiving antenna with a gain of 4.5dBi. The Beidou module is equipped with the Hexin Xingtong UM982, which supports B1I / B3I dual-frequency positioning with a positioning accuracy of up to 2.5m and a timing accuracy of 100ns. It is equipped with a 26MHz high-stability TCXO clock with a clock stability of ±0.5ppm and an Allan variance of 1e-11@1s, providing a highly stable clock signal for the operation of each module.

[0154] The computing and storage module undertakes the core tasks of local model inference, algorithm execution, data storage and scheduling on the terminal, while taking into account the rational allocation of computing power output and storage resources. The AI ​​accelerator uses Huawei Ascend 310, which can achieve a computing power output of 16 TOPS (INT8) with a power consumption of only 8W, adapting to the terminal's local lightweight inference needs. The main control AP processor is LC1860C, which is responsible for application layer algorithm execution and human-computer interaction. The storage configuration includes a 4GB eMMC chip and an eMCP chip. The 4GB eMMC chip is managed by functional partitions, with the 500MB system area set to read-only mode, the 100MB track data area, the 50MB model area, and the 3GB historical data area using a circular overwrite mode. The eMCP chip is a combination configuration of 2G (DDR3) + 16G (eMMC) to meet the storage needs of various data and programs on the terminal.

[0155] The structure and power supply module adopt an industrial-grade rugged design, adapting to the deployment and working requirements of various harsh outdoor environments; the shell is made of aluminum alloy, achieving an IP67 protection rating, and is equipped with a heat dissipation fin design to ensure the device's heat dissipation efficiency; the power supply system supports a wide voltage input of DC10V~36V, has a built-in 100Wh lithium battery, and supports power hot switching to avoid power interruption affecting device operation; the device can operate stably in a wide temperature range of -40℃ to +70℃, adapting to application scenarios in different regions and climates.

[0156] The core process for terminal beam optimization in general scenarios such as open areas or urban areas mainly includes four stages: hardware initialization and data synchronization, parallel acquisition of multi-source data, data preprocessing and feature calculation, and feature normalization and matrix construction.

[0157] After the terminal powers on, it first completes the initialization and data synchronization of the core hardware modules, providing a basic guarantee for subsequent beam optimization:

[0158] Module initialization: Start the Beidou dual-frequency positioning and timing module, RF front-end and AI chip, and complete the register configuration and function self-test of each module;

[0159] Time synchronization: The Beidou module output signal and UTC time message are synchronized to the baseband processing unit and the RTC of the main control MCU through hardware interrupts to ensure that the time deviation of the entire system is less than 1 millisecond;

[0160] Ephemeris Loading: The simplified broadcast ephemeris of the Tiantong-1 satellite for the next 24 hours is read from the orbital data partition of the local storage. This ephemeris contains a table of orbital elements and perturbation correction coefficients every 15 minutes, providing data support for subsequent orbital feature calculations.

[0161] During the synchronization period, the terminal performs three types of tasks in parallel: radio frequency measurement, BeiDou positioning, and spectrum sensing, to complete the acquisition of multi-dimensional raw data.

[0162] RF measurement task: Control the RF front end to quickly scan all preset beam center frequencies (7 beams in the example) within the downlink S-band (2170-2200MHz) of Tiantong-1; collect 100ms of IQ data at each beam frequency point, and calculate the original received signal strength (RSSI_raw[i]), original signal-to-noise ratio (SNR_raw[i]), and root mean square delay spread (τ_rms[i]) of the beam in real time.

[0163] BeiDou positioning task: Receive standard NMEA-0183GNGGA (positioning information) and GNZDA (time information) messages output by the BeiDou module through the UART serial port, parse and obtain the terminal latitude (Lat), longitude (Lon), altitude (Alt), UTC time (t_UTC), and positioning validity flag; if the positioning is valid, synchronously read the carrier phase observation value in the module register for subsequent high-precision time synchronization verification;

[0164] Spectrum sensing task: Using the spectrum analysis mode of AD9361, perform a 1024-point FFT scan on the 2170-2200MHz band to identify and record the center frequency and power I_peak(f) of all interference peaks that exceed the background noise level by more than 10dB.

[0165] The collected raw data is preprocessed to calculate core features such as trajectory, interference, and position.

[0166] Spatiotemporal alignment: All measurement data are stamped with a unified BeiDou timestamp t_k to ensure data temporal consistency;

[0167] Signal filtering: Apply Kalman filtering to the RSSI_raw[i] sequence (state equation is x_t=x_{t-1}, observation equation is z_t=x_t+v_t, process noise and observation noise variance are preset according to open / urban scene), and output smoothed RSSI_Kalman; calculate the variance σ²_SNR of the most recent 10 samples for the SNR_raw[i] sequence;

[0168] Doppler calculation: The carrier phase change rate is calculated by using the difference between two adjacent 1-second period carrier phase observations, and then the radial velocity and Doppler frequency offset f_Doppler relative to the satellite are obtained (formula: f_Doppler=(Δφ / (2π))*f_c, where f_c is the carrier frequency and Δφ is the differential carrier phase after deambiguation).

[0169] Orbital characteristic calculation: Based on t_k and satellite ephemeris, the satellite's position [X_s,Y_s,Z_s] in the geocentric inertial frame is calculated. The terminal WGS-84 coordinates are converted to geocentric inertial frame coordinates [X_u,Y_u,Z_u]. The angular position of the satellite in the orbital plane is calculated and normalized to [0,2π) to obtain the basic phase. Perturbation correction is completed by combining the preset perturbation table, and finally the instantaneous phase angle θ(t_k) of the satellite is obtained. At the same time, the satellite position is numerically differentiated to calculate the rate of change of satellite-to-ground distance dr / dt and the rate of change of orbital eccentricity de / dt.

[0170] Interference characteristics calculation: I_in-band is obtained by integrating the interference power spectral density within ±1MHz of the current beam frequency. The signal-to-interference-plus-noise ratio (SINR) is estimated by combining RSSI_Kalman and the system thermal noise floor.

[0171] Location feature calculation: The Geohash algorithm is used to encode the latitude and longitude of the terminal into a 32-bit integer GridHash, and the altitude is normalized to the [0,1] interval to obtain Alt_norm. The elevation angle of the satellite relative to the terminal is calculated. The number of successful connections of beams in the same grid, similar phase and time range is queried from the local SQLite historical database to calculate the beam preference index BeamPref∈[0,1].

[0172] The feature normalization process is completed by performing linear normalization on the original feature values ​​based on the pre-stored historical maximum (max) and minimum (min) values ​​of each feature dimension: F_norm[i]=(F_raw[i]-min[i]) / (max[i]-min[i]). Example: The RSSI_Kalmin range is preset to [-130dBm,-60dBm]. If the current value is -85dBm, the normalized value is (-85+130) / (70)≈0.643.

[0173] When constructing the input feature matrix of the deep learning model, the 16 feature values ​​are stored in a float32 array F

[16] in a fixed order, namely: [RSSI_Kalman,σ²_SNR,f_Doppler,σ_θ,A_24h,A_6h,dSNR / dt,θ(t),dr / dt,de / dt,I_in-band,SINR_est,GridHash(int->float),Alt_norm,Elevation,BeamPref]. This forms a multi-dimensional feature vector for the model input.

[0174] For indoor repeater scenarios, a special beam optimization process is added to ensure scenario adaptability:

[0175] When the data collected by the terminal meets at least three of the following conditions, it is determined to be a repeater scenario:

[0176] RSSI > -90dBm and SNR < 8dB; RSSI / SNR ratio exceeds 3 times the typical value in open environment;

[0177] The spectrum exhibits narrowband interference peaks at specific frequency points such as 1.8MHz and 3.6MHz; the multipath delay spread τ_rms > 40ns;

[0178] The signal envelope exhibits a "flat-top" distortion characteristic.

[0179] To address the interference characteristics in a repeater scenario, the useful signal component and the interference noise component introduced by the repeater are first separated:

[0180] Signal decomposition: the process of analyzing the pre-processed received signal. The model is constructed, assuming it consists of three parts: the satellite's useful signal. Interference noise introduced by repeaters and receiver thermal noise .Right now: = + + .

[0181] Interference Feature Extraction and Separation: Taking advantage of the strong temporal correlation and relatively fixed spectral characteristics of repeater interference (such as broadband noise floor elevation and specific intermodulation products), blind source separation techniques such as eigenvalue decomposition (EVD) or independent component analysis (ICA) are employed to analyze the covariance matrix of the received signal. By identifying components orthogonal to or with minimal correlation to the satellite signal steering vector space, the interference subspace is estimated, and thus, the interference features are separated. Component estimates .

[0182] Power estimation:

[0183] Calculate the total power of the received signal: dt. Calculate the power of the separated interference noise: dt. Receiver thermal noise power These are known hardware parameters.

[0184] Calculate the true signal-to-noise ratio According to the definitions of signal and noise, the useful signal power can be approximated as: However, for the sake of conservatism and stability, a more robust calculation method is to treat repeater interference and thermal noise as the total "severe noise," as shown in the following formula:

[0185] ;

[0186] The essence of this formula is to subtract the identifiable repeater interference power from the total power, treat the remaining power as the power that may contain the signal, and then compare it with the total noise power (interference + thermal noise). This method can effectively eliminate the power inflation caused by the repeater's amplification effect and more realistically reflect the communication potential of the satellite link.

[0187] Next, interference separation processing is performed: using the known spreading pilot sequence, the ideal channel response h_ideal(t) is estimated through a correlation receiver; the total channel response h_actual(t) is estimated from the actual received signal; the residual h_noise(t) = h_actual(t) - h_ideal(t) is calculated to characterize the interference and non-ideal channel components; FFT analysis is performed on the residual h_noise(t) to analyze the spectral components, and the energy overlapping with the repeater's operating frequency band and known interference frequencies is marked as the interference component. Based on the interference separation results, the signal purity is calculated, and the beamforming score weights are adjusted.

[0188] Signal purity is defined as: =(‖h_ideal‖²+‖h_multipath‖²) / ‖h_actual‖²; where h_multipath is the natural multipath component. The interference threshold in the dynamic weight calculation formula is adjusted based on the purity value. :

[0189] <0.3 (Severe Distortion): Setting =-90dBm;

[0190] 0.3≤ <0.6 (moderate distortion): Set =-85dBm;

[0191] ≥0.6 (Quality acceptable): Setting =-80dBm;

[0192] Signal purity The focus is on measuring the proportion of interference, while It is a key metric directly used for communication link budget. The two can be used in combination, for example, when... When it is too low, even if the calculated value is... It is acceptable, but the actual bit error rate may deteriorate due to the non-Gaussian nature of the interference. In this case, the score of the beam needs to be reduced by adjusting the weights.

[0193] This embodiment uses a hybrid architecture model deployed on the Ascend 310 chip as an example to illustrate the engineering implementation process of local inference and beam decision:

[0194] Model deployment and loading:

[0195] Using the Huawei MindSpore framework, the trained CNN-BiLSTM-Transformer hybrid model is pruned and quantized, then converted into the .om offline model format supported by Ascend 310, with the model size controlled within 48MB. When the terminal starts, the beam_selection_model.om model is loaded into the chip's DDR memory through the Ascend Runtime API, the input and output tensors are initialized, and preparations for inference are completed.

[0196] Model inference process (spatial feature branch):

[0197] The 16-dimensional feature vector F

[16] is copied to the input memory buffer of Ascend 310 and reshaped into a tensor format of [1,1,16] (single sample, sequence length 1, 16-dimensional features), which meets the requirements of one-dimensional convolution input.

[0198] Conv1D-1: Configures 32 convolutional kernels of length 3, performs one-dimensional convolution in the feature dimension, outputs a tensor of [1,1,32], followed by ReLU activation and BatchNorm layers;

[0199] Conv1D-2 / 3: Convolutional kernels of length 5 and 7 are used sequentially, with output channels of 64 and 128 respectively, gradually fusing interactive information from different feature dimensions.

[0200] BiLSTM layer execution: Input the [1,1,128] tensor output by Conv1D-3 into a bidirectional LSTM layer with 64 hidden units, and take the output h_t

[128] of the last time step;

[0201] Spatial score output: h_t is passed through a fully connected layer W_s[128x1] and a bias b_s to obtain a scalar s_raw, which is then mapped to the [0,1] interval by the Sigmoid function to output the spatial feature score S_spatial.

[0202] Model inference process (time feature branch)

[0203] Position encoding: Pre-compute the position encoding matrix PE (dimension d_model=16), map the satellite phase angle θ(t) to the position index pos, and add PE[pos] to the input feature F element by element to obtain F'=F+PE[pos];

[0204] Transformer encoder forward propagation:

[0205] Linear projection: F' generates a query (Q), key (K), and value (V) vector with dimension d_k=4 through three linear layers;

[0206] Multi-head attention: Q, K, and V are split into 4 heads, and scaled dot product attention is calculated for each head. The outputs of each head are concatenated and passed through a linear layer to obtain Z_attn.

[0207] Residual connection and layer normalization: Z1=LayerNorm(F'+Z_attn), after two layers of MLP (hidden layer dimension 64, ReLU activation), residual connection and normalization are performed again to obtain the final output Z

[16] ;

[0208] Temporal score output: The Z vector is processed through two layers of MLP (16→8→1) and Sigmoid activation to output the temporal feature score. ∈[0,1].

[0209] Dynamic weight calculation and decision fusion (executed by the main control AP processor)

[0210] Read the current beamband interference power from the AD9361 register. (dBm), and calculate the measurement uncertainty based on the variance of multiple measurements. ; Based on the scene type (open area / city / repeater station), read the preset parameters from the ROM and substitute them into the formula to calculate the base weight α_base; if it is a repeater station scene, then... Value adjustment dynamic weight (Non-repeater scenario) =α_base); query the exponential attenuation weighted success rate of candidate beams over the past 7 days from the historical database. ;

[0211] Overall score calculation: Formula: ;

[0212] Select The beam corresponding to the maximum value is the optimal beam. The control radio front end switches to the center frequency of this beam and initiates a channel access request. The timestamp, grid hash, phase θ, and values ​​of each beam in this decision are then used to make the decision. The optimal beam ID and successful service establishment flag are written to the historical database circular buffer.

[0213] See Figure 4 Federated learning update sequence diagram, relying on satellite channels to achieve federated learning updates, the specific process is as follows:

[0214] Periodic triggering and local training

[0215] The terminal checks local historical data every 30 days. If there are more than 1,000 valid records, it starts local training: loads the triplet data of the past 180 days (feature F, selected beam b, service success flag); uses cross-entropy as the loss function and adopts the SGD optimizer (lr=1e-5, momentum 0.9) to fine-tune the model for 3-5 rounds; calculates the gradient update amount ΔW=W_local-W_global_old.

[0216] Gradient compression and protection (compressed data <50KB)

[0217] Multi-level compression and privacy protection are performed on gradient update amounts:

[0218] Sparsification: Retain the top 1% of gradient values ​​with the largest absolute value in ΔW;

[0219] Non-uniform quantization: 8-bit quantization, with a fine step size of 0.001 for small gradients and a coarse step size of 0.01 for large gradients;

[0220] Differential coding: The difference between the code and the underlying gradient table;

[0221] Entropy coding: Huffman coding further compresses the code;

[0222] Differential privacy: Adding Gaussian noise .

[0223] Satellite uplink transmission

[0224] Receive update window instructions broadcast by the Tiantong network (example: UTC02:00-04:00); monitor channel quality (SNR, bit error rate), and adaptively adjust transmission parameters (use 8PSK2 / 3 code rate when SNR>10dB, otherwise use QPSK1 / 2 code rate); adopt ARQ mechanism and breakpoint resume strategy, assign a unique sequence number to each gradient packet, and upload it to the ground station through the satellite channel.

[0225] Ground station security aggregation

[0226] Receive the terminal gradient packet and complete decryption and digital signature verification;

[0227] Anomaly detection: Discard anomaly updates with gradient norms exceeding the mean ± 3σ;

[0228] Weighted FedAvg aggregation: Assign weights according to "terminal data volume × channel quality factor" and calculate the global gradient ΔW_global;

[0229] Generate global update: .

[0230] Satellite broadcast delivery and terminal updates

[0231] The ground station will cyclically transmit W_global_new through the Tiantong-1 broadcast channel (for example, repeating every 5 minutes for 2 hours). The packet structure includes version number, hash value, model parameters, and CRC32 check bits.

[0232] After receiving the data, the terminal verifies the version number increment and hash consistency, writes the new parameters to the Flash spare area, completes the model hot update during business intervals, and switches to the new model for operation.

[0233] Closed-loop evolution

[0234] The decision data generated by the new model is continuously stored in the local historical database, serving as the data source for local training in the next 30-day cycle. This forms a closed-loop evolutionary system of "data collection - model inference - gradient update - global aggregation - model iteration," enabling continuous optimization of model performance.

[0235] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0236] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of this disclosure.

[0237] 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 disclosure 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.

[0238] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for dynamic beam optimization in satellite communication, characterized in that, The terminal's locally pre-built hybrid architecture deep learning model also includes the following steps: The terminal collects spatiotemporal characteristic data of its environment in real time, including spatial characteristics, temporal characteristics, orbital characteristics, interference characteristics and location characteristics, and stores them in a local historical database; Based on the current location features, timestamp, and satellite ephemeris, the terminal calculates the satellite's instantaneous phase angle, identifies its own scene type, and executes appropriate signal processing strategies for the interference features according to different scenes to obtain optimized interference features; the formula for calculating the satellite's instantaneous phase angle is as follows: in, This is the current timestamp. This serves as a reference time for the start of the orbital period. To reference the cumulative number of leap seconds from the previous epoch to the present, For the satellite orbital period, For the initial phase shift, The instantaneous phase angle, The instantaneous orbital angle deviation caused by the perturbation; The orbital features, spatial features, temporal features, and positional features of multiple historical frames that match the current satellite's instantaneous phase angle are extracted from the local historical database. These features are then combined with the optimized interference features to construct a multi-dimensional feature matrix. The multi-dimensional feature matrix is ​​then input into the deep learning model to obtain spatial feature scores and temporal feature scores. Dynamic weights are calculated by combining real-time interference levels with the scene type. The decision-making proportions of spatial feature scores and temporal feature scores are balanced by the dynamic weights. The comprehensive score of each beam is then calculated by integrating historical reputation scores. The beam with the highest comprehensive score is selected as the optimal beam. The local decision-making results are recorded in the local historical database for subsequent calculation of the historical reputation scores of each beam. The dynamic weight Calculated using the following adaptive Sigmoid function: ； in, To measure the interference power, To mitigate measurement uncertainty, For scene-related interference thresholds, Let k be the inherent uncertainty of the system, and k be the slope factor of the Sigmoid function. The formula for calculating the overall score is as follows: in, For dynamic weights, Scoring the spatial characteristics of beam b Scoring the time characteristics of beam b Historical reputation decay factor Score the historical reputation of beam b.

2. The method as described in claim 1, characterized in that, The scenario type includes at least an indoor repeater scenario, and the corresponding signal processing strategy includes: performing interference separation processing to separate the noise introduced by the indoor repeater from the natural multipath interference, and calculating the true signal-to-noise ratio and signal purity; updating the interference features with the true signal-to-noise ratio as the core parameter, and using the signal purity to adjust the interference threshold of the function for calculating the dynamic weight.

3. The method as described in claim 2, characterized in that, When the scenario is identified as an indoor repeater, the interference threshold Based on the signal purity calculated in this scenario The adjustments will be made dynamically, and the specific rules are as follows: when When <0.3, set =-90dBm; When 0.3≤ When <0.6, set =-85dBm; when When ≥0.6, set = -80dBm.

4. The satellite communication beam dynamic optimization method as described in claim 1, characterized in that, The hybrid architecture deep learning model includes a spatial feature extraction part and a temporal feature extraction part: The spatial feature extraction part adopts a combination structure of convolutional neural network (CNN) and long short-term memory network (BiLSTM) to extract spatial correlation features and output spatial feature scores. The time feature extraction part adopts a Transformer network structure to extract time series features and output time feature scores.

5. The satellite communication beam dynamic optimization method as described in claim 1, characterized in that, The deep learning model is updated using a satellite channel-based federated learning mechanism, including: The terminal performs model fine-tuning based on data from the local historical database, generates gradient update data, performs multi-level compression processing on the gradient update data, and uploads the multi-level compressed gradient update data to the satellite ground station via the satellite channel. The satellite ground station securely aggregates gradient update data from multiple terminals to generate a global model update, which is then broadcast and distributed via satellite channel. The terminal receives and applies the global model update.

6. The satellite communication beam dynamic optimization method as described in claim 5, characterized in that, The multi-level compression includes gradient sparsity processing, non-uniform integer quantization, differential coding, and entropy coding compression; the secure aggregation includes first filtering the gradient data for abnormal gradients to remove invalid and abnormal gradient information, then performing the aggregation operation by adding differential privacy noise and using homomorphic encryption, and simultaneously performing weighted aggregation based on the local data quality and data volume of each terminal.

7. A satellite communication terminal, characterized in that, To implement the satellite communication beam dynamic optimization method according to claim 1, the method includes: The spatiotemporal feature acquisition module is used to collect spatiotemporal feature data of the surrounding environment in real time, including spatial features, temporal features, orbital features, interference features, and location features. The feature construction module calculates the instantaneous phase angle of the satellite based on the current position features, timestamp, and satellite ephemeris, identifies the scene type, and performs appropriate signal processing strategies for the interference features for different scenes to obtain optimized interference features; it extracts the orbit features, spatial features, temporal features, and position features of historical multiple frames that match the current instantaneous phase angle of the satellite from the local historical database, and constructs a multi-dimensional feature matrix by combining the optimized interference features; The beam scoring selection module inputs the multi-dimensional feature matrix into the hybrid architecture deep learning model to perform local inference, and outputs spatial feature scores and temporal feature scores respectively; it calculates dynamic weights by combining real-time interference level and scene type, and integrates historical reputation scores to obtain a comprehensive score for each beam; finally, it selects the beam with the highest comprehensive score as the optimal beam, and records the local decision results to the local historical database for subsequent calculation of historical reputation scores for each beam. A local historical database is used to store the spatiotemporal feature data and to store a pre-built deep learning model with a hybrid architecture.

Images