Low earth orbit satellite beam adaptive acquisition method based on fusion double architecture

Abstract

This invention relates to the field of satellite communication technology, specifically to a low-Earth orbit (LEO) satellite beam adaptive acquisition method based on a fused dual-architecture approach. The method includes: achieving local clock calibration and acquiring the almanac of the target LEO satellite by locking onto the broadcast channel of a high-Earth orbit satellite; utilizing an inertial measurement unit (IMU) and a temperature sensor network to collect terminal pose and array temperature gradient in real time, constructing a multi-dimensional physical feature tensor. Subsequently, the feature tensor is input into a nonlinear compensation model integrating multi-scale dilated convolution and spatial self-attention mechanisms, and a phase predistortion matrix is ​​output. By performing a channel-level Hadamard product operation between the predistortion matrix and the theoretical pointing matrix, a beam control driving matrix is ​​generated, and a physical receiving beam is synthesized. After capturing the downlink signal, the true Doppler frequency shift residual is extracted, and an error autocorrelation matrix is ​​constructed to inversely correct the almanac compensation weights. This invention achieves robust beam acquisition through multi-dimensional feature perception and a closed-loop algorithmic logic.

Description

Technical Field

[0001] This invention relates to the field of satellite communication technology, specifically to a low-Earth orbit satellite beam adaptive acquisition method based on a fused dual architecture. Background Technology

[0002] With the rapid advancement of global low-Earth orbit (LEO) satellite internet construction, portable satellite communication terminals play a crucial role in achieving seamless global coverage and high-bandwidth mobile communication services. In LEO satellite communication systems, the extremely high orbital speeds of LEO satellites relative to ground terminals typically lead to significant Doppler shifts and rapid, real-time changes in the signal incident angle. For portable terminals, achieving stable signal access within a limited visual window requires precise carrier frequency synchronization, code phase acquisition, and antenna beam alignment under complex dynamic conditions. In existing signal processing workflows, terminals typically need to perform initial pointing based on pre-set satellite almanac data and use attitude sensors to detect terminal motion data for real-time compensation to maintain antenna beam alignment with the satellite. However, during actual deployment and movement, the acquisition performance of portable terminals is often affected by the coupling interference of multi-dimensional nonlinear factors. For example, fluctuations in high-frequency hardware processes within the terminal, drastic changes in ambient temperature causing phase drift in the RF channel, and beam distortion and gain attenuation occurring when the phased array antenna performs large-angle off-axis scanning—these factors collectively constitute a complex error environment.

[0003] In summary, how to improve the acquisition accuracy and adaptive calibration capability of low-Earth orbit satellite communication beams under multi-dimensional coupling conditions such as high dynamic frequency shift of satellites, terminal environmental interference, and nonlinear distortion of underlying hardware, so as to ensure the rapid establishment and continuous stability of communication links, is a technical problem that urgently needs to be solved by those skilled in the art.

[0004] To address this, a low-orbit satellite beam adaptive acquisition method based on a fused dual architecture is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive beam acquisition method for low-Earth orbit (LEO) satellites based on a fused dual-architecture approach. Through multi-dimensional feature perception and closed-loop algorithmic logic, robust beam acquisition is achieved. This includes: local clock calibration and acquisition of the target LEO satellite's almanac by locking onto the LEO satellite's broadcast channel; real-time acquisition of terminal pose and array temperature gradient using an inertial measurement unit and a temperature sensor network to construct a multi-dimensional physical feature tensor; inputting the feature tensor into a nonlinear compensation model integrating multi-scale dilated convolution and spatial self-attention mechanisms to infer and output a phase predistortion matrix; generating a beam control drive matrix and synthesizing the physical receiving beam by performing a channel-level Hadamard product operation between the predistortion matrix and the theoretical pointing matrix; and extracting the true Doppler frequency shift residual after acquiring the downlink signal to construct an error autocorrelation matrix and inversely correcting the almanac compensation weights.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A low-Earth orbit satellite beam adaptive acquisition method based on a fusion dual-architecture architecture is applied to a portable station terminal with a built-in dual-system baseband module, including: The almanac data and global reference clock of the target low-orbit satellite are extracted through the high-orbit communication link of the dual-system baseband module; the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna are collected. The local reference time is aligned using a global reference clock. Based on the local reference time, almanac data, and three-dimensional pose data, the initial line-of-sight angle and theoretical Doppler shift of the target low-orbit satellite relative to the phased array antenna are calculated, and the wide-angle off-axis angle is separated. Read the pre-stored fingerprint error matrix, generate a phase predistortion matrix based on the wide-angle off-axis angle, operating temperature and fingerprint error matrix; generate a theoretical pointing matrix based on the initial viewing angle, and perform a Hadamard product operation between the theoretical pointing matrix and the phase predistortion matrix to generate a wave control drive matrix; The beam control drive matrix is ​​input into the radio frequency front end of the phased array antenna to generate a physical receiving beam. Through the low-orbit communication link of the dual-mode baseband module, the downlink synchronization signal of the target low-orbit satellite is received using the physical receiving beam, and the downlink synchronization signal is analyzed. Extract the true Doppler frequency shift from the downlink synchronization signal, calculate the error autocorrelation matrix between the true Doppler frequency shift and the theoretical Doppler frequency shift, and inject the error autocorrelation matrix into the dual-mode baseband module to update the compensation weights of the ephemeris data.

[0007] Preferably, the built-in dual-mode baseband module includes a high-orbit baseband processing unit, a low-orbit baseband processing unit, and a data interaction bus; the high-orbit communication link is a narrowband signaling control link established based on the high-orbit baseband processing unit, used to maintain the underlying handshake with the auxiliary high-orbit satellite in a low signal-to-noise ratio environment; the low-orbit communication link is a broadband service data link established based on the low-orbit baseband processing unit; the data interaction bus is used to transmit the almanac data and global reference clock parsed by the high-orbit baseband processing unit to the low-orbit baseband processing unit in real time as a priori compensation input for low-orbit signal acquisition.

[0008] Preferably, the process of acquiring data via the high-orbit communication link through the dual-system baseband module includes: locking the downlink broadcast control channel of the auxiliary high-orbit satellite, demodulating the downlink broadcast control channel, parsing the orbital parameters of the target low-orbit satellite and the global reference clock signal used to calibrate the local oscillator, and using them as the almanac data; the acquisition of the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna includes: measuring the pitch angle, roll angle, and yaw angle of the portable station terminal in real time through an inertial measurement unit rigidly connected to the phased array antenna to form the three-dimensional pose data; and acquiring the operating temperature data characterizing the thermal drift state of the radio frequency channel in real time through a temperature sensor network mounted on the bottom layer of the phased array antenna.

[0009] Preferably, the process of calculating the initial line-of-sight angle and theoretical Doppler shift of the target low-Earth orbit satellite relative to the phased array antenna, and separating the wide-angle off-axis angle, includes: calibrating the local controlled oscillator inside the portable station terminal using the global reference clock signal to eliminate hardware clock bias and output the local reference time; calculating the real-time spatial position vector and real-time velocity vector of the target low-Earth orbit satellite in a preset geocentric coordinate system based on the local reference time and the almanac data; constructing the local carrier coordinate system of the phased array antenna according to the three-dimensional pose data; and projecting the real-time spatial position vector onto the local carrier coordinate system using a coordinate transformation matrix to obtain... A relative position vector is obtained, and the elevation and azimuth components of the relative position vector are extracted to form the initial line-of-sight angle. The real-time velocity vector is transformed to the local carrier coordinate system to obtain the relative velocity vector. The radial projection component of the relative velocity vector in the direction of the initial line-of-sight angle is calculated, and the theoretical Doppler frequency shift is generated based on the radial projection component and the communication carrier frequency. The true spatial angle between the initial line-of-sight angle and the physical normal vector of the phased array antenna is calculated. When it is determined that the true spatial angle is greater than a preset scanning gain attenuation threshold, the true spatial angle is extracted and separated as the wide-angle off-axis angle.

[0010] Preferably, the process of generating the waveguide drive matrix includes: concatenating the wide-angle off-axis angle, the operating temperature data, and the fingerprint error matrix into a multi-dimensional physical feature vector; inputting the multi-dimensional physical feature vector into a pre-trained nonlinear array compensation model to output the phase predistortion matrix for the phased array antenna; calculating the ideal spatial path difference of each RF channel based on the physical element distribution topology of the phased array antenna, combined with the azimuth and elevation components of the initial line of sight, and generating the theoretical pointing matrix containing ideal phase weights; extracting the ideal phase elements of each channel in the theoretical pointing matrix, and the corresponding conjugate compensation elements in the phase predistortion matrix, performing the Hadamard product operation on the corresponding elements, and generating the waveguide drive matrix containing comprehensive calibration weights by channel-level mapping multiplication.

[0011] Preferably, the nonlinear array compensation model includes a cascaded tensor construction layer, a multi-scale dilated convolutional layer, a spatial self-attention layer, and a feature tiling layer: the tensor construction layer reads the fingerprint error matrix and reconstructs it into an initial feature tensor consistent with the two-dimensional physical arrangement dimension of the phased array antenna; the initial feature tensor is input to the multi-scale dilated convolutional layer, and multiple sets of convolutional kernels with different dilation rates are used to perform feature extraction in parallel to obtain a spatial electromagnetic coupling feature map between near-field adjacent array elements and far-field cross-region array elements inside the phased array antenna; the spatial electromagnetic coupling feature map is input to the spatial self-attention layer, the dot product correlation matrix between the feature vectors of each array element channel is calculated, and a spatial attention weight matrix is ​​generated based on the dot product correlation matrix; the spatial electromagnetic coupling feature map and the spatial attention weight matrix are multiplied element-wise, and a dimension reduction and flattening operation is performed through the feature tiling layer to output a physical coupling feature map, and the phase predistortion matrix is ​​determined based on the physical coupling feature map.

[0012] Preferably, the process of parsing the downlink synchronization signal includes: extracting the complex weight elements in the wave control drive matrix and mapping them down to each RF transceiver channel of the phased array antenna RF front end; configuring the feed phase and attenuation amplitude of the numerically controlled phase shifter and variable attenuator inside the corresponding RF transceiver channel using the complex weight elements, and synthesizing the physical receiving beam pointing to the initial line of sight and the controlled sidelobe level through coherent superposition of space electromagnetic waves; capturing space electromagnetic signals using the physical receiving beam, performing low-noise amplification, down-conversion and analog-to-digital conversion operations on the space electromagnetic signals, and inputting the baseband digital signal into the low-Earth orbit communication link of the dual-system baseband module; performing sliding matched filtering with the baseband digital signal using the pre-stored local synchronization sequence in the low-Earth orbit communication link to extract the relevant peak energy; when the relevant peak energy is continuously detected to cross the preset acquisition decision threshold, confirming successful reception of the downlink synchronization signal of the target low-Earth orbit satellite, and parsing and extracting the physical layer frame number and timing synchronization identifier in the downlink synchronization signal.

[0013] Preferably, the process of updating the compensation weights of the almanac data includes: performing carrier frequency offset estimation and phase-locked loop residual extraction on the downlink synchronization signal to obtain frequency offset observations over multiple consecutive symbol periods to form the time series of the true Doppler frequency shift; performing a difference operation between the time series of the true Doppler frequency shift and the theoretical Doppler frequency shift to generate a time-domain Doppler error vector; performing an outer product operation between the time-domain Doppler error vector and its transpose to generate an error autocorrelation matrix; using the error autocorrelation matrix as the dynamic observation noise covariance, and injecting it inversely into the orbit tracking filter of the high-orbit baseband processing unit of the dual-system baseband module; using the state update equation of the orbit tracking filter to extract the space orbit perturbation compensation amount based on the eigenvalues ​​of the error autocorrelation matrix; using the space orbit perturbation compensation amount to update the local propagation model parameters of the almanac data to generate the updated compensation weights, so as to adaptively calibrate the initial line-of-sight angle in the next acquisition period.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By pre-extracting almanac data and a global reference clock through the high-orbit communication link, "spatiotemporally guided" acquisition is achieved during the low-orbit satellite access process. Using the long-period stable signaling of the high-orbit satellite as a reference benchmark, the search space of the low-orbit beam is compressed from a blind search across the entire domain to a narrow-domain pointing based on the initial line-of-sight angle. This dual-architecture collaborative mechanism not only reduces the computing power overhead of the portable station terminal in the initial access phase, but also ensures that the terminal can quickly rebuild the link based on the high-orbit prior information in complex weather or instantaneous obstruction environments.

[0015] 2. By introducing a fingerprint error matrix and a nonlinear array compensation model, adaptive correction of phase distortion caused by process fluctuations and thermal drift in the sixteen-layer board lamination process was achieved. This ensures that the phased array antenna can still produce a high-gain, low-sidelobe physical beam when scanning at large off-axis angles. This not only reduces the reliance of portable stations on expensive high-frequency boards and extreme processing precision, but also effectively suppresses beam distortion and energy diffusion during wide-angle scanning.

[0016] 3. By extracting the true Doppler frequency shift of the low-Earth orbit downlink signal and injecting it in reverse into the dual-mode baseband module, a closed-loop feedback loop from "measured physical characteristics" to "prior ephemeris model" is constructed, achieving deep fusion between heterogeneous data. This scheme utilizes the frequency offset residuals measured from the low-Earth orbit link to generate an error autocorrelation matrix, which can identify and correct the orbital parameter offsets of the ephemeris transmitted from the high-Earth orbit link in real time, thereby adaptively calibrating the compensation weights in subsequent acquisition periods. This closed-loop mechanism solves the prediction bias problem caused by low-Earth orbit satellite orbit perturbations. Attached Figure Description

[0017] Figure 1 A schematic diagram of the low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture provided by the present invention; Figure 2 This is a schematic diagram of the process for generating the wave control drive matrix according to the present invention; Figure 3 This is a schematic diagram of the compensation weight process for updating calendar data according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1 to 3 This invention provides a low-Earth orbit satellite beam adaptive acquisition method based on a fused dual architecture, the technical solution of which is as follows: A low-Earth orbit satellite beam adaptive acquisition method based on a fusion dual-architecture approach is applied to a portable station terminal with a built-in dual-system baseband module, and the process is as follows: Figure 1 As shown, it includes: The almanac data and global reference clock of the target low-orbit satellite are extracted through the high-orbit communication link of the dual-system baseband module; the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna are collected. The local reference time is aligned using a global reference clock. Based on the local reference time, almanac data, and three-dimensional pose data, the initial line-of-sight angle and theoretical Doppler shift of the target low-orbit satellite relative to the phased array antenna are calculated, and the wide-angle off-axis angle is separated. Read the pre-stored fingerprint error matrix, generate a phase predistortion matrix based on the wide-angle off-axis angle, operating temperature and fingerprint error matrix; generate a theoretical pointing matrix based on the initial viewing angle, and perform a Hadamard product operation between the theoretical pointing matrix and the phase predistortion matrix to generate a wave control drive matrix; The beam control drive matrix is ​​input into the radio frequency front end of the phased array antenna to generate a physical receiving beam. Through the low-orbit communication link of the dual-mode baseband module, the downlink synchronization signal of the target low-orbit satellite is received using the physical receiving beam, and the downlink synchronization signal is analyzed. Extract the true Doppler frequency shift from the downlink synchronization signal, calculate the error autocorrelation matrix between the true Doppler frequency shift and the theoretical Doppler frequency shift, and inject the error autocorrelation matrix into the dual-mode baseband module to update the compensation weights of the ephemeris data.

[0020] Example 1: Furthermore, the built-in dual-mode baseband module includes a high-orbit baseband processing unit, a low-orbit baseband processing unit, and a data interaction bus; the high-orbit communication link is a narrowband signaling control link established based on the high-orbit baseband processing unit, used to maintain the underlying handshake with the auxiliary high-orbit satellite in a low signal-to-noise ratio environment; the low-orbit communication link is a broadband service data link established based on the low-orbit baseband processing unit; the data interaction bus is used to transmit the almanac data and global reference clock parsed by the high-orbit baseband processing unit to the low-orbit baseband processing unit in real time as a priori compensation input for low-orbit signal acquisition.

[0021] Specifically, the built-in dual-mode baseband module is deployed on the physical hardware in a heterogeneous computing platform that includes a field-programmable gate array (FPGA) and a digital signal processing unit (DSP). At the logic resource level inside the FPGA, two independent hardware processing areas are divided through logic partitioning technology, which are configured as a high-track baseband processing unit and a low-track baseband processing unit, respectively. Between these two logic processing areas, a shared dual-port random access memory (DPRAM) and an AXI interconnect bus are configured to jointly form a data interaction bus.

[0022] During the terminal device startup phase, the high-orbit baseband processing unit allocates the underlying narrowband digital downconverter and long coherent integrator accumulator. After the RF front-end performs analog-to-digital conversion on the received high-orbit satellite frequency band signal, the high-orbit baseband processing unit performs carrier stripping and spreading code sliding correlation operations on the digital signal. When the detected correlation peak energy crosses the decision threshold, the underlying control logic locks the carrier phase, completing the physical layer handshake of the narrowband signaling control link. Subsequently, the DSP depackets and decodes the demodulated signaling data frame, extracts the latest almanac data segment of the target low-orbit satellite from the frame payload, and simultaneously extracts the global reference clock data containing the absolute timestamp. The DSP converts the extracted global reference clock data into a high-orbit time synchronization control command and sends it to the FPGA logic layer. Based on this control command, the high-orbit baseband processing unit inside the FPGA, combined with the locked high-orbit spreading code phase boundary and the local crystal counter, accurately recovers the 1PPS pulse signal aligned with the system time through hardware logic.

[0023] During the cross-system data transfer process, the DSP extracts the calendar data and absolute timestamps and writes them into the designated physical address space of the shared DPRAM on the data interaction bus according to a preset structure format. After the write operation is completed, the data interaction bus sends a high-level hardware interrupt signal to the low-rail baseband processing unit through the internal interrupt controller. At the same time, the 1PPS pulse signal generated by the high-rail baseband processing unit is directly hardwired to the local timer reset port of the low-rail baseband processing unit through the low-latency dedicated clock routing network configured inside the FPGA.

[0024] Upon receiving a hardware interrupt signal and a 1PPS pulse signal, the LEO baseband processing unit triggers the clearing of the local clock register and the loading of the absolute timestamp, completing the time alignment at the baseband level. Subsequently, the LEO baseband processing unit reads almanac data from the shared DPRAM via the AXI interconnect bus, combines it with the currently aligned time reference, obtains the local three-dimensional coordinates and velocity vectors pre-configured by the terminal itself or input through an external navigation module (such as an inertial navigation / GNSS module), calls the SGP4 (Simplified Conventional Perturbation) orbit prediction model, and performs orbital position extrapolation and relative radial velocity calculation for the target LEO satellite. Specifically, it calculates the relative position vector between the current spatial position vector of the LEO satellite and the terminal's local three-dimensional coordinates, and projects the three-dimensional velocity vector of the LEO satellite onto the radial direction of the relative position vector to obtain the relative radial velocity.

[0025] The low-Earth orbit baseband processing unit (LEO) calculates the expected Doppler frequency shift value, combines it with the local system clock frequency of the LEO and the phase accumulator bit width (e.g., 32-bit or 48-bit) of the numerically controlled oscillator (NCO), calculates the ratio of the expected Doppler frequency shift to the system clock frequency, multiplies this ratio by two raised to the power of the phase accumulator bit width, performs full-scale mapping, and after rounding and quantization, converts it into a binary fixed-point number format frequency control word, which is directly injected into the internal NCO. The NCO then generates a local oscillator signal with reverse frequency offset compensation based on this frequency control word. When establishing a broadband service data link, the low-Earth orbit (LEO) baseband processing unit uses the pre-compensated local oscillator signal to mix with the broadband LEO satellite signal input from the radio frequency front end to eliminate carrier frequency offset caused by high-speed motion. Since the Doppler change rate is extremely large when the LEO satellite passes overhead, the LEO baseband processing unit activates a timer interrupt mechanism to repeatedly execute the above relative radial velocity calculation according to a millisecond-level update cycle, and refreshes the frequency control word of the numerically controlled oscillator (NCO) in real time to dynamically track and compensate for large-slope carrier frequency offset changes. Subsequently, it drives the broadband matched filter bank to perform pseudo-random sequence correlation detection on the mixed baseband data to complete the acquisition of the LEO signal.

[0026] By locking the high-orbit broadcast control channel and analyzing the global clock, combined with an inertial measurement unit rigidly connected to the phased array antenna and an underlying temperature sensor network, a high-precision absolute physical reference is provided for local oscillator calibration and spatial coordinate dimensionality reduction. This design accurately captures environmental parameters that truly characterize the thermal drift state of the underlying RF channel, effectively ensuring the reliability of data input for subsequent Doppler frequency shift calculation and RF hardware distortion compensation.

[0027] Furthermore, the process of acquiring data via the high-orbit communication link through the dual-system baseband module includes: locking the downlink broadcast control channel of the auxiliary high-orbit satellite, demodulating the downlink broadcast control channel, parsing the orbital parameters of the target low-orbit satellite and the global reference clock signal used to calibrate the local oscillator, and using them as the almanac data; the acquisition of the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna includes: measuring the pitch angle, roll angle, and yaw angle of the portable station terminal in real time through an inertial measurement unit rigidly connected to the phased array antenna to form the three-dimensional pose data; and acquiring the operating temperature data characterizing the thermal drift state of the radio frequency channel in real time through a temperature sensor network mounted on the bottom layer of the phased array antenna.

[0028] When the terminal is powered on or idle, the RF transceiver front end of the portable station terminal tunes its local oscillator frequency to the downlink communication band of the auxiliary high-orbit satellite. The high-orbit RF signal received by the antenna is amplified by low noise and down-converted by RF, and then input to the analog-to-digital converter to be converted into a digital baseband signal. The digital matched filter in the high-orbit baseband processing unit uses the locally pre-stored broadcast control channel synchronization sequence to perform a sliding correlation operation on the digital baseband signal. When a correlation peak is detected, the carrier recovery loop is triggered to lock the signal carrier and drive the demodulator to complete the signal demodulation. Subsequently, the frame parsing logic extracts the frame header delimiter from the demodulated bit stream according to the frame structure protocol of the high-orbit communication system, locates and strips the signaling payload data. The baseband processing unit parses the six roots or ephemeris parameters of the target low-orbit satellite from the payload data and simultaneously extracts the global reference clock signal packet with the system absolute timestamp. These two parts of data are reassembled together into the almanac data for subsequent calculations.

[0029] For the parsed global reference clock signal, the baseband processing unit compares the absolute timestamp contained within it with the local system time maintained by the internal counter of the locally controlled crystal oscillator to obtain the current clock phase deviation value. This clock phase deviation value is smoothed by an internal digital loop filter and converted into a voltage control signal or a digital frequency control word. The underlying hardware control logic injects this control command into the adjustment port of the locally controlled crystal oscillator. By dynamically fine-tuning the output frequency of the crystal oscillator, the accumulated frequency drift and phase difference between the local clock and the high-orbit global reference clock are eliminated, completing the hardware-level calibration of the local oscillator. Specifically, the baseband processing unit uses the synchronization pulse generated during the demodulation of the high-orbit signal as a hardware trigger signal. At the moment the pulse arrives, the hardware logic circuit synchronously latches the value of the local counter. The latched value is then aligned and subtracted from the absolute timestamp to eliminate the non-deterministic delay fluctuations caused by interrupt response and software processing logic.

[0030] In the attitude data acquisition stage, the inertial measurement unit (IMU) inside the terminal is physically anchored to the printed circuit board substrate of the phased array antenna via rigid supports or direct surface-mount welding, ensuring strict physical alignment of their three-dimensional spatial coordinate systems. Specifically, the installation error matrix of the IMU coordinate system relative to the phased array antenna electromagnetic beam coordinate system is pre-obtained through turntable calibration. During the calculation process, this installation error matrix is ​​used to perform coordinate rotation transformation on the original pose data, ensuring that the compensated pose data accurately reflects the spatial orientation of the antenna beam axis. The three-axis gyroscope and three-axis accelerometer inside the IMU continuously acquire data on the terminal's position in space at a fixed sampling rate. The three-axis angular velocity and three-axis accelerometer data are read through a serial peripheral interface bus and the attitude fusion solution logic is executed. During the solution process, the gravity vector direction measured in real time by the accelerometer is used as a reference. The low-frequency cumulative drift error generated by the integration operation of the raw gyroscope data is continuously corrected by Kalman filtering. Finally, the pitch angle, roll angle and yaw angle, which represent the current real space pointing of the terminal antenna array, are calculated and output to form three-dimensional pose data. The inertial measurement unit also includes a three-axis magnetometer, which collects the geomagnetic vector and combines it with the gravity vector provided by the accelerometer to provide a geomagnetic north reference for the attitude fusion solution logic, thereby correcting the integral drift of the gyroscope in the yaw angle.

[0031] In the temperature data acquisition stage, multiple digital temperature sensors are mounted in a grid pattern on the feed network and backplane area of ​​the RF transceiver component at the bottom of the phased array antenna, forming a temperature sensor network. The physical location of each temperature sensor corresponds to a RF power amplifier or the core chip of the transmit / receive component with high heat dissipation characteristics. The system environment monitor of the terminal reads the digital temperature value output by each node in the grid sequentially according to a preset polling cycle via a multiplexed bus. The environment monitor stitches and aligns all the collected discrete temperature data according to the geometric arrangement topology of the physical array elements, reconstructing a two-dimensional data table that records the real-time temperature gradient of the antenna array surface. This data is used as the operating temperature data characterizing the current physical thermal drift state of each RF channel and input into the shared memory. The reconstruction process uses the Kriging interpolation algorithm, with the measured values ​​of the temperature sensor network as nodes, to calculate the estimated temperature value of the area on the array surface without sensor coverage, thereby generating a continuous temperature gradient distribution map covering the corresponding physical locations of all RF channels, realizing a refined characterization of the thermal drift state of different channels.

[0032] By establishing a high-precision time base and multi-dimensional spatial coordinate transformation, the complex relative motion between satellite and ground is converted into accurate line-of-sight angle and frequency offset prediction values. This effectively eliminates the interference of local hardware clock bias on acquisition accuracy and utilizes off-axis angle quantization to identify the critical region of beam gain attenuation.

[0033] Furthermore, the process of calculating the initial line-of-sight angle and theoretical Doppler shift of the target low-Earth orbit satellite relative to the phased array antenna, and separating the wide-angle off-axis angle, includes: calibrating the local controlled oscillator inside the portable station terminal using the global reference clock signal to eliminate hardware clock bias and output the local reference time; calculating the real-time spatial position vector and real-time velocity vector of the target low-Earth orbit satellite in a preset geocentric coordinate system based on the local reference time and the almanac data; constructing the local carrier coordinate system of the phased array antenna based on the three-dimensional pose data, and projecting the real-time spatial position vector into the local carrier coordinate system using a coordinate transformation matrix. Obtain the relative position vector and extract its elevation and azimuth components to form the initial line-of-sight angle; transform the real-time velocity vector to the local carrier coordinate system to obtain the relative velocity vector, calculate the radial projection component of the relative velocity vector in the direction of the initial line-of-sight angle, and calculate the theoretical Doppler frequency shift based on the radial projection component and the communication carrier frequency; calculate the true spatial angle between the initial line-of-sight angle and the physical normal vector of the phased array antenna, and when the true spatial angle is determined to be greater than a preset scanning gain attenuation threshold, extract and separate the true spatial angle as the wide-angle off-axis angle.

[0034] After receiving a global reference clock signal (such as a second pulse signal containing an absolute timestamp) from a high-orbit satellite, the terminal initiates a local timing synchronization task. This task latches and compares the value of a high-frequency counter driven by a local controlled oscillator with the externally input absolute timestamp, calculates the time deviation between the two, and then converts the deviation value into a frequency adjustment command through a digital proportional-integral (DPI) filtering algorithm. This command is then applied to the frequency control terminal of the local controlled oscillator. The execution period of the DPI algorithm is set to match the synchronization period of the second pulse signal, ensuring that a command update is triggered every time a pulse arrives. Simultaneously, the step size of the frequency adjustment command is quantized and smoothed according to the frequency modulation sensitivity of the local controlled oscillator to prevent instantaneous step jumps in the local oscillator frequency, and outputs a local reference time for subsequent orbit calculations.

[0035] A pre-set satellite orbit prediction model (such as the SGP4 model) is invoked, using the Kepler orbital elements from the almanac data and the local reference time as input parameters. Through recursive orbital dynamics calculations, the three-dimensional spatial position vector and three-dimensional velocity vector of the target low-Earth orbit satellite relative to a pre-set geocentric coordinate system (such as the ECEF coordinate system) at the current instant are calculated. This process transforms the originally static almanac parameters into six-dimensional spatial state quantities characterizing the satellite's dynamic operational state, serving as the raw input for subsequent coordinate transformations.

[0036] In the process of constructing the local carrier coordinate system, the pitch, roll, and heading angles output by the inertial measurement unit are read. Combined with the terminal's geographical location information (latitude, longitude, and altitude), a rotation transformation matrix from the geocentric coordinate system to the phased array antenna's local carrier coordinate system is calculated and generated. Specifically, the calculation and generation process is divided into two steps: First, based on the terminal's geographical location information, a station-centric East-North-Sky coordinate system with the terminal's location as the origin is established, and the transformation matrix from the geocentric coordinate system to the station-centric coordinate system is calculated. Then, using the three-dimensional pose data, the attitude rotation matrix from the station-centric coordinate system to the local carrier coordinate system based on the antenna array surface is calculated. By multiplying the above two matrices, the final rotation transformation matrix is ​​obtained. The transformation matrix contains three-dimensional Euler angle rotation components and a translation component determined by geographical location. Using this matrix, the real-time spatial position vector of the target low-orbit satellite in the geocentric coordinate system is transformed by coordinate projection to a local carrier coordinate system with the geometric center of the phased array antenna as the origin, thereby obtaining the relative position vector of the target satellite relative to the antenna array. The ratio of the horizontal and vertical components of this relative position vector is further extracted. By calculating the arctangent and arcsine functions, the initial line of sight angle of the satellite relative to the antenna array, namely the azimuth and elevation angles, is obtained.

[0037] To calculate the theoretical Doppler shift, the real-time velocity vector of the target low-Earth orbit satellite is transformed to the local carrier coordinate system using the rotation transformation matrix, obtaining the relative velocity vector of the target satellite relative to the terminal. When calculating this relative velocity vector, the terminal's own real-time motion velocity vector (e.g., the three-dimensional ground velocity obtained through the BeiDou / GPS navigation module) needs to be acquired simultaneously. Then, in the local carrier coordinate system, the terminal's own real-time motion velocity vector is subtracted from the target low-Earth orbit satellite's real-time velocity vector to obtain the complete relative velocity vector. Subsequently, the projection component of this relative velocity vector onto the direction vector pointed to by the initial line-of-sight angle, i.e., the radial velocity component, is calculated. The ratio of this radial velocity component to the speed of light is used as a scaling factor and multiplied by the nominal carrier frequency of the satellite communication system to generate the theoretical Doppler shift value caused by the satellite's high-speed dynamic operation. This frequency shift value is transmitted in real-time to the baseband processing module to pre-compensate for carrier frequency offset during signal acquisition.

[0038] During the separation phase of the wide-angle off-axis angle, the physical normal vector of the phased array antenna surface (usually defined as a unit vector perpendicular to the surface) is obtained. The dot product between the unit vector of the direction represented by the initial line of sight and the physical normal vector is calculated. The true spatial angle between the satellite line of sight and the antenna normal is obtained by calculating the inverse cosine of the dot product. The true spatial angle is compared with the scanning gain attenuation threshold (such as the gain drop point caused by the antenna beam scanning to a specific angle) stored in the memory. When it is determined that the true spatial angle exceeds the threshold, the angle is marked and extracted as the wide-angle off-axis angle, which serves as the criterion for triggering subsequent nonlinear model compensation. The scanning gain attenuation threshold is determined by measuring the gain envelope curve of the phased array antenna at different scanning angles in advance in a microwave anechoic chamber. The scanning angle corresponding to the point where the beam gain drops to 3 dB of the center value is defined as the edge critical angle, and the threshold is set according to the edge critical angle.

[0039] By introducing a fingerprint error matrix and combining it with multi-dimensional physical features for nonlinear compensation, refined calibration of antenna manufacturing deviations, thermal drift, and wide-angle scanning distortion was achieved. The pre-distortion matrix was precisely mapped to the RF channel level using the Hadamard product, effectively correcting pointing deviations.

[0040] Further, the process of generating the waveguide drive matrix includes: concatenating the wide-angle off-axis angle, the operating temperature data, and the fingerprint error matrix into a multi-dimensional physical feature vector; inputting the multi-dimensional physical feature vector into a pre-trained nonlinear array compensation model to output the phase predistortion matrix for the phased array antenna; calculating the ideal spatial path difference of each RF channel based on the physical element distribution topology of the phased array antenna, combined with the azimuth and elevation components of the initial line of sight, and generating the theoretical pointing matrix containing ideal phase weights; extracting the ideal phase elements of each channel in the theoretical pointing matrix, and the corresponding conjugate compensation elements in the phase predistortion matrix, performing the Hadamard product operation on the corresponding elements, and generating the waveguide drive matrix containing comprehensive calibration weights by channel-level mapping multiplication, the specific process is as follows: Figure 2 As shown.

[0041] First, the separated wide-angle off-axis angle data and real-time acquired operating temperature data are retrieved from the memory buffer. Then, the fingerprint error matrix pre-stored for the current phased array antenna is read from the non-volatile memory. According to the preset data structure, the discrete angle scalar, temperature scalar, and error matrix containing the static phase deviation values ​​of each channel are dimensionally aligned and concatenated to construct a unified multi-dimensional physical feature vector. This vector, as a comprehensive feature set including spatial geometry, environmental thermals, and inherent hardware characteristics, is cached in the model input buffer. Specifically, the dimensional alignment includes: expanding the fingerprint error matrix in row-major order according to the array element index to transform it into a one-dimensional error feature vector; and appending the wide-angle off-axis angle and operating temperature data to the end of the one-dimensional error feature vector after normalizing them with gain coefficients to form a fixed-length linear input vector.

[0042] Subsequently, the multidimensional physical feature vector is input into a pre-trained nonlinear array compensation model integrated within the baseband chip. This model consists of multiple fully connected or convolutional layers, and its internal weights have been trained using electromagnetic simulation data and actual RF channel measurement data in an offline environment. Model inference calculations are then performed, and the input vector is linearly combined and mapped using a nonlinear activation function across different weight layers. This nonlinear activation function uses a hyperbolic tangent function to map the hidden layer features to the interval -1 to 1. The output layer uses a linear scaling layer to map the interval values ​​to -1. to The radian value between them is used as the original value of phase compensation for the corresponding RF channel to ensure that the model output matches the control range of the physical phase shifter. Finally, a two-dimensional matrix containing phase compensation values ​​corresponding one-to-one with the number of physical array elements of the phased array antenna is reconstructed in the output layer of the model. This matrix is ​​called the phase predistortion matrix. Each element in the matrix represents the amount of conjugate phase correction that needs to be superimposed to compensate for RF channel distortion at a specific angle and temperature.

[0043] In the generation stage of the theoretical pointing matrix, the physical element distribution topology data of the phased array antenna, which is pre-stored in the system configuration file, is obtained. This data records the precise three-dimensional coordinates of each radiating element in the array relative to the geometric center of the array surface. Combined with the previously calculated azimuth and elevation components of the initial line of sight, a spatial vector projection algorithm is used to calculate the geometric path difference of the target plane wave relative to the reference origin when it reaches the position of each physical element. These physical spatial path differences are then divided by the wavelength of the current communication carrier and multiplied by twice pi to convert them into corresponding electrical phase offset values. After conversion to these electrical phase offset values, a modulo operation of twice pi is performed to reduce the phase values ​​of all channels to a closed interval from zero to twice pi, in order to conform to the physical operating characteristics of the numerically controlled phase shifter. These ideal phase values ​​calculated for all RF channels are arranged in a matrix form consistent with the element arrangement, constituting the theoretical pointing matrix.

[0044] Finally, the matrix operation acceleration unit is activated to perform channel-level phase fusion operations. Simultaneously, the ideal phase elements corresponding to each position in the theoretical pointing matrix and the conjugate compensation elements at the same coordinate index in the phase predistortion matrix are extracted. The calculation unit performs a Hadamard product operation on these two sets of elements, that is, in the complex or polar coordinate domain, the ideal pointing phase and the predistortion compensation phase are added element-wise (or multiplied by the corresponding complex mapping). Through this channel-level one-to-one mapping operation, the theoretical pointing command and the hardware nonlinear calibration quantity are fused into a set of comprehensive phase weights, which are then encapsulated into the final wave control drive matrix. This matrix is ​​subsequently converted into control signals to drive the digital phase shifters of each channel in the RF front-end. The conversion process includes: obtaining the quantization bit depth of the digital phase shifter; dividing the phase interval from zero to two times π into corresponding discrete steps according to the quantization bit depth; mapping each element in the comprehensive phase weights to the nearest discrete step index to generate the corresponding binary wave control codeword; and encapsulating the wave control codeword into a wave control command frame according to the RF front-end bus protocol.

[0045] By constructing a tensor reconstruction and multi-scale dilated convolution architecture in the hardware topology-aware branch, a refined modeling of the physical arrangement of phased array elements and the spatial electromagnetic coupling effect is achieved. This design utilizes the multi-scale receptive field of dilated convolution to simultaneously extract intra-array near-field interference and cross-regional far-field coupling features, and dynamically quantizes the interference weights between array elements through a spatial self-attention mechanism, thereby transforming scattered hardware errors into feature maps with physical topological properties.

[0046] Furthermore, the nonlinear array compensation model includes a cascaded tensor construction layer, a multi-scale dilated convolutional layer, a spatial self-attention layer, and a feature tiling layer: the tensor construction layer reads the fingerprint error matrix and reconstructs it into an initial feature tensor consistent with the two-dimensional physical arrangement dimension of the phased array antenna; the initial feature tensor is input to the multi-scale dilated convolutional layer, and multiple sets of convolutional kernels with different dilation rates are used to perform feature extraction in parallel to obtain the spatial electromagnetic coupling feature map between near-field adjacent array elements and far-field cross-region array elements inside the phased array antenna; the spatial electromagnetic coupling feature map is input to the spatial self-attention layer, the dot product correlation matrix between the feature vectors of each array element channel is calculated, and a spatial attention weight matrix is ​​generated based on the dot product correlation matrix; the spatial electromagnetic coupling feature map is multiplied element-wise with the spatial attention weight matrix, and the feature tiling layer performs dimensionality reduction and flattening operations to output the physical coupling feature map.

[0047] At the initial stage of model inference, the tensor construction layer reads the one-dimensional fingerprint error matrix data from the underlying memory-mapped region. Based on the actual physical array element arrangement parameters of the phased array antenna (such as the number of horizontal and vertical array elements), it performs resampling and dimension reconstruction on the one-dimensional data. Specifically, this layer expands the complex phase deviation of each channel into two independent components, real and imaginary, and fills them into the corresponding two-dimensional grid coordinates, thereby constructing an initial feature tensor with three-dimensional depth. This process logically realizes the transformation from an abstract electrical signal vector to a physical topological tensor with spatial orientation attributes. Specifically, the dimension reconstruction adopts a row-first mapping rule consistent with the physical arrangement of the phased array antenna, filling the complex deviation value of the i-th array element into the row and column coordinates corresponding to the tensor. At the same time, in order to ensure the local perception effectiveness of the convolution operation, zero-padding is performed on the edges of the two-dimensional grid to maintain the consistency between the output tensor and the input array surface size.

[0048] The initial feature tensor is then processed in parallel by a multi-scale dilated convolutional layer. This layer is configured with multiple convolutional kernel branches with different dilation rates. The first branch uses a convolutional kernel with a dilation rate of one to extract the direct mutual coupling features between adjacent array elements through a sliding window. The second and higher-order branches use dilation rates greater than one to skip adjacent grids and capture the secondary electromagnetic induction features between far-field array elements over a larger spatial span. The local feature maps extracted by each branch are concatenated and stitched together in the channel dimension to form a spatial electromagnetic mutual coupling feature map that integrates multi-scale spatial correlations. This expands the receptive field of the model for the overall error evolution of the array without increasing the number of computational parameters. To ensure dimensional alignment during stitching, the multiple convolutional kernel branches all use the same convolutional kernel size (e.g., 3×3) and unit stride. During parallel execution, branches with dilation rates greater than one expand the sampling interval by inserting null points between kernel elements, thereby maintaining the consistency of feature map resolution at the output and facilitating direct concatenation and stitching in the channel dimension.

[0049] The spatial self-attention layer receives the aforementioned mutually coupled feature maps and divides them into feature vector blocks corresponding to multiple physical array elements. Before performing linear transformation, each feature vector block needs to be superimposed with a preset two-dimensional spatial position code. The position code is generated based on the physical coordinates of the array elements on the phased array surface. By injecting spatial position information into the feature vectors, subsequent dot product operations can identify the proximity correlation characteristics between array elements in different physical regions. This layer maps each feature vector block into a query vector, a key vector, and a value vector through a linear transformation matrix. It calculates the dot product between the query vector and the key vector corresponding to any two spatial positions to generate a correlation score matrix characterizing the degree of mutual interference among all array elements. Subsequently, amplitude normalization processing is performed on this score matrix to generate a spatial attention weight matrix. Specifically, the normalization processing uses a normalized exponential function to map the energy value after the dot product calculation to the interval between zero and one, thereby mathematically characterizing the relative contribution intensity of each array element feature to the final pointing distortion. This weighting matrix accurately characterizes the contribution of array element errors at different spatial locations to the overall beam distortion at the current operating frequency and angle, achieving nonlinear enhancement of key interference features.

[0050] In the final processing stage, the feature tiling layer performs element-level weighted fusion and dimensionality reduction operations. This layer multiplies the spatial electromagnetic mutual coupling feature map with the spatial attention weight matrix at corresponding spatial locations point by point, so that the model attention is focused on the mutual coupling region that contributes significantly to the pointing deviation. The weighted feature tensor is then compressed into a one-dimensional physical mutual coupling feature map through global average pooling or direct flattening logic. This feature map contains the hard core error features after spatial topology calibration, which serves as the core feature input for the subsequent output phase predistortion matrix of the model, ensuring that the compensation result is highly coupled with the physical and electromagnetic characteristics of the antenna array.

[0051] Before the model is deployed in practice, a training dataset covering multiple physical conditions needs to be constructed through a combination of offline simulation and field calibration. First, a full-wave electromagnetic model of the phased array antenna is established using electromagnetic simulation software (such as HFSS or CST). Various process deviation parameters are preset in the model, including substrate dielectric constant fluctuations, stack thickness tolerances, and mutual coupling coefficient matrices between array elements. By using parametric scanning, the frequency sweep range of the simulated signal, the spatial incident angle (azimuth and elevation), and the feed amplitude noise are changed to generate a large number of mapping relationship samples between the ideal pointing phase and the actual distorted beam, which serve as the initial simulation dataset.

[0052] Subsequently, during the hardware mass production phase, actual calibration data acquisition was performed. The portable station terminal integrating the phased array antenna was placed in a microwave anechoic chamber, and the actual beam pattern at different scanning angles was measured using a far-field probe. During the test, the ambient temperature of the terminal was changed by an external temperature control device, and the static phase deviation and gain fluctuation of each RF channel at different temperature points were recorded and used as a fingerprint error matrix. For each measured operating point, an optimization algorithm (such as the rotating vector method or amplitude-phase joint optimization) was used to obtain the ideal phase calibration vector that maximizes the current beam gain and minimizes the pointing error. This vector was used as the ground truth label under the input feature, thereby constructing a measured dataset containing the real nonlinear characteristics of the hardware. In the optimization process, the ratio coefficient of the pointing error weight and the gain fluctuation weight was preset, and the beam main lobe pointing deviation was less than a preset percentage as a mandatory constraint. The gradient search algorithm was used to find the phase combination that maximizes the overall array synthesis gain in the phase space, thereby ensuring the physical and logical uniqueness of the generated ground truth label.

[0053] In the dataset preprocessing stage, the simulation data and measured data are fused together, and data augmentation is performed. Before being packaged into training samples, the multidimensional physical feature vectors are normalized preprocessed, and the wide-angle off-axis angle, operating temperature data, and fingerprint error matrix are mapped to a dimensionless interval between zero and one to eliminate the interference of scale differences between different physical dimensions on the model gradient update. Specifically, the fusion process adopts a transfer learning strategy. First, the model is pre-trained using the large initial simulation dataset to enable the model to master the basic electromagnetic propagation physical characteristics. Then, fine-tuning training is performed using the smaller but more accurate measured dataset, and the loss weight of the measured samples is set higher than that of the simulation samples to force the model to prioritize fitting the nonlinear characteristics of real hardware. Specifically, random phase perturbations conforming to a Gaussian distribution are injected into the fingerprint error matrix to simulate the aging process of electronic components. At the same time, non-uniform resampling is performed on the wide-angle off-axis angle to increase the sample weight in the large-angle scanning region (gain attenuation region). Finally, each processed data set is encapsulated into a training sample pair containing the input feature vector and the ideal pre-distortion truth position label, and divided into training set, validation set and test set according to the proportion, and stored in the training database of the high-performance computing platform.

[0054] The training process of the model is executed on a server with GPU acceleration capabilities. First, the weight matrices of each layer inside the model are randomly initialized (such as using Kaiming initialization or Xavier initialization). In each training iteration, the training engine extracts a batch of training samples from the database and inputs them into the network architecture consisting of tensor construction layers, multi-scale dilated convolutional layers, spatial self-attention layers and feature tiling layers, performs forward propagation calculations, and outputs the predicted phase predistortion matrix.

[0055] Subsequently, the training engine calls the loss function calculation unit to perform a difference operation between the predicted phase output by the model and the ideal phase in the sample label. The loss function adopts the mean square error function and introduces a penalty factor including a Laplacian regularization term to suppress drastic changes between the compensation weights of each channel and ensure beam smoothness. Specifically, when calculating the phase difference, a periodic mapping logic is introduced to constrain the difference between the model's predicted phase and the ideal label phase within the range of positive and negative pi. The square of the difference of the shortest path is taken as the input of the loss function to conform to the circular continuity characteristic of the phase physical quantity. The Laplacian regularization term is achieved by calculating the second-order discrete difference between the weight of each radio frequency channel and the weights of its four adjacent channels in the two-dimensional physical arrangement. By penalizing the phase mutation value, the model output is forced to continuously change the compensation matrix in the spatial topology, avoiding spatial phase noise caused by over-compensation of individual channels. The calculated loss value is updated layer by layer by the backpropagation algorithm according to the direction of gradient descent to update the parameters of each convolution kernel and attention weight of the model.

[0056] The training process employs an adaptive learning rate optimization algorithm (such as the Adam optimizer). By monitoring the loss decline curve on the validation set, the convergence step size is dynamically adjusted. When the model's prediction accuracy on the validation set reaches a preset threshold (e.g., the root mean square of the phase deviation is less than a preset degree), and the loss no longer decreases significantly for several consecutive cycles, the iteration stops. At this point, the trained model weight parameters are exported and compressed using fixed-point quantization, packaged into a binary model file, and burned into the baseband processing chip of the portable terminal for real-time inference. The fixed-point quantization includes: statistically analyzing the dynamic distribution range of the output values ​​of each layer of the model and determining the fixed-point truncation scaling factor accordingly; using asymmetric quantization to map floating-point parameters to a preset bit-width (e.g., eight-bit or sixteen-bit) integer format; and compensating for phase pointing offset caused by rounding errors through quantization-aware training or precision recalibration on the validation set.

[0057] Further, the process of parsing the downlink synchronization signal includes: extracting the complex weight elements in the wave control drive matrix and mapping them down to each RF transceiver channel of the phased array antenna RF front end; configuring the feed phase and attenuation amplitude of the numerically controlled phase shifter and variable attenuator inside the corresponding RF transceiver channel using the complex weight elements, and synthesizing the physical receiving beam with a specific spatial pointing and sidelobe level through coherent superposition of space electromagnetic waves; capturing space electromagnetic signals using the physical receiving beam, performing low-noise amplification, down-conversion and analog-to-digital conversion operations on the space electromagnetic signals, and inputting the baseband digital signal into the low-Earth orbit communication link of the dual-system baseband module; in the low-Earth orbit communication link, performing sliding matched filtering with the baseband digital signal using a pre-stored local synchronization sequence to extract the relevant peak energy; when the relevant peak energy is continuously detected to cross the preset acquisition decision threshold, confirming successful reception of the downlink synchronization signal of the target low-Earth orbit satellite, and parsing and extracting the physical layer frame number and timing synchronization identifier in the downlink synchronization signal.

[0058] First, the generated wave control drive matrix in memory is accessed. This matrix is ​​stored as a complex array. The real and imaginary parts of each matrix element are extracted and mapped to digital control words with specific widths according to the RF channel control protocol. Specifically, the mapping process includes: using a coordinate rotation digital calculation algorithm or a lookup table method to convert the complex weights into corresponding phase angles and amplitude attenuation factors; then, based on the bit width of the digitally controlled phase shifter (e.g., 6-bit quantization) and the bit width of the variable attenuator (e.g., 5-bit quantization), the phase angle and attenuation factor are linearly mapped to corresponding binary index values, and the gain and phase step size are corrected using a pre-stored RF channel nonlinear calibration table, thereby generating the final digital control words. Subsequently, these digital control words are distributed to various RF transceiver components of the phased array antenna RF front-end through a high-speed serial peripheral interface or parallel bus. Within each RF channel, the control word is received and converted into a drive voltage or current signal, which is applied to the control terminals of the digitally controlled phase shifter and the variable attenuator, thereby precisely adjusting the initial feed phase and amplitude attenuation of the channel signal. The electromagnetic waves radiated by each array element coherently superimpose in space according to Huygens' principle, forming phase-aligned constructive interference at the target pointing position, thereby synthesizing a physical receiving beam with a preset main lobe pointing and controlled side lobe level.

[0059] The spatial electromagnetic signal captured by the physical receiving beam enters the receiving link of the RF front-end. First, it undergoes weak signal amplification through a low-noise amplifier to suppress noise contributions from subsequent circuits. The amplified signal is then fed into a mixer, where it undergoes down-conversion with the local oscillator signal, shifting the high-frequency RF signal to the intermediate frequency or baseband frequency range. Subsequently, the signal passes through a low-pass or band-pass filter to remove image interference and out-of-band noise, and then enters an analog-to-digital converter for high-speed sampling, converting it into a digital baseband signal stream. To ensure the accuracy of subsequent sliding matching, the sampling frequency of the analog-to-digital converter is set to be greater than twice the signal symbol rate to achieve fractional oversampling. The sampled signal undergoes multi-rate conversion through a digital decimation filter, ensuring that the sampling clock of the output digital baseband signal stream is synchronized with the local processing master frequency. This digital signal stream is input in real-time to the hardware logic area responsible for low-orbit service processing in the dual-mode baseband module via a parallel differential bus (such as LVDS) or a high-speed serial protocol.

[0060] At the digital processing front end of the low-Earth orbit communication link, a preset local synchronization sequence is invoked. Before performing sliding matched filtering, the previously calculated theoretical Doppler frequency shift is extracted, and a complex rotator (such as the CORDIC operator) is driven to perform reverse phase compensation on the baseband digital signal to ensure that the signal energy can be effectively aggregated during the matched filtering process. This sequence is consistent with the synchronization pilot characteristics defined by the target satellite protocol. A sliding matched filter based on a finite impulse response structure is constructed. The input baseband digital signal is sequentially pushed into the shift register sequence and point-by-point complex multiplication and addition operations are performed with the local synchronization sequence. The result of each moment output by the filter is subjected to modulus square operation to extract its envelope energy. This process aggregates the dispersed signal energy into a correlation peak energy pulse with extremely high signal-to-noise ratio through continuous sliding matched calculation in the time domain.

[0061] The energy sequence output by the matched filter is continuously monitored and compared with a preset acquisition decision threshold. The acquisition decision threshold is generated in real time by a constant false alarm rate detection algorithm. The average noise power of the non-correlated area within the current observation window is statistically analyzed and multiplied by a preset threshold coefficient to form a dynamic threshold, thereby achieving robust acquisition of the synchronization signal under different weather and interference environments. To prevent false triggering caused by noise bursts, the decision unit adopts a continuous confirmation mechanism. That is, when the energy of the correlation peak is detected to cross the threshold in multiple consecutive preset observation windows or symbol periods, it is determined that the acquisition logic is locked. Once the lock is successful, the downlink synchronization signal is time-aligned using the instantaneous position of the correlation peak as the time zero point, and the physical layer decoding engine is driven to descramble and decode the subsequent broadcast channel data blocks. Finally, the physical layer frame number representing the system time series and the timing synchronization identifier used to determine the cell boundary are extracted from the demodulated original bit stream.

[0062] During the system design phase, the preset threshold coefficient is determined through the following steps: First, based on the physical layer protocol requirements of the low-Earth orbit satellite communication system, the maximum allowable false alarm probability is determined (e.g., set to 10 to the power of -6), which is the upper limit of the probability that the decision unit incorrectly determines a successful acquisition in the presence of only noise. Since the envelope energy of the matched filter output follows a Rayleigh or chi-square distribution in a pure noise environment, the theoretical multiple required to maintain this false alarm rate is derived backward based on the logarithmic function relationship of the false alarm probability. Subsequently, a full-system link simulation is performed in a microwave anechoic chamber or simulated channel environment. By injecting additive white Gaussian noise of different intensities, the ratio distribution of the correlation peak energy to the average noise floor under different signal-to-noise ratio conditions is statistically analyzed. Based on the detection probability curve (ROC curve) obtained from the simulation, the above theoretical multiple is fine-tuned and calibrated while ensuring a high acquisition probability (e.g., greater than 99%). Finally, the calibrated value is used as the threshold coefficient and stored in a non-volatile register in a fixed-point format. During the actual acquisition process, the value of this register is directly read and multiplied by the average noise power calculated in real time. In addition, to cope with complex electromagnetic interference environments, the coefficient can be preset in multiple levels according to the current communication mode. For example, a higher threshold coefficient can be used in the initial blind search phase to suppress false alarms, and a lower threshold coefficient can be used in the stable tracking phase to prevent link dropout.

[0063] By correlating the actual Doppler frequency shift obtained from demodulation with theoretical predictions, a closed-loop feedback path was constructed from the signal processing output to the front-end orbital dynamics model. This allows the system to dynamically capture the minute residuals between prior almanac parameters, local crystal oscillator temperature drift, and the actual satellite operating state, and to use the statistical properties of the error autocorrelation matrix to correct the compensation weights online. This mechanism not only effectively suppresses the cumulative error in orbit prediction over time, but also narrows the time-frequency search window in the subsequent acquisition process through a continuously converging feedback loop.

[0064] Further, the process of updating the compensation weights of the almanac data includes: performing carrier frequency offset estimation and phase-locked loop residual extraction on the downlink synchronization signal to obtain frequency offset observations over multiple consecutive symbol periods to form the time series of the true Doppler frequency shift; performing a difference operation between the time series of the true Doppler frequency shift and the theoretical Doppler frequency shift to generate a time-domain Doppler error vector; performing an outer product operation between the time-domain Doppler error vector and its transpose to generate an error autocorrelation matrix; using the error autocorrelation matrix as the dynamic observation noise covariance, and back-injecting it into the orbit tracking filter of the high-orbit baseband processing unit of the dual-system baseband module; using the state update equation of the orbit tracking filter, extracting the space orbit perturbation compensation amount based on the eigenvalues ​​of the error autocorrelation matrix; using the space orbit perturbation compensation amount to update the local propagation model parameters of the almanac data, generating the updated compensation weights to adaptively calibrate the initial line-of-sight angle in the next acquisition period, the specific process is as follows: Figure 3 As shown.

[0065] Once the downlink synchronization signal is locked, the carrier tracking loop enters a stable tracking state. The internal digital phase-locked loop and frequency-locked loop continuously perform phase and frequency discrimination operations on the received signal. The accumulated output value of the loop filter in the carrier recovery loop and the instantaneous error output of the frequency discriminator are extracted in real time. These values ​​are sampled at fixed time steps to obtain a digital sequence representing the carrier frequency shift. Subsequently, mean filtering or low-pass filtering is applied to this sequence to remove high-frequency jitter caused by thermal noise, ultimately yielding a series of observation samples representing the true Doppler frequency shift, constituting a time series of the true Doppler frequency shift.

[0066] The acquired actual Doppler shift time series was synchronized with the theoretical Doppler shift previously calculated using the orbital model. Subtraction was performed to calculate the deviation between the observed and theoretical values ​​at each moment, thereby extracting the Doppler error, which includes orbital element perturbations, residual temperature drift of the local crystal oscillator, and nonlinear distortion of the propagation path. These discrete deviation values ​​were arranged in chronological order to construct a time-domain Doppler error vector. This vector physically reflects the mismatch between the prior almanac model and the actual physical link state.

[0067] Subsequently, the time-domain Doppler error vector is multiplied by its transpose. In this calculation, each element of the vector is multiplied by itself and all other elements to generate a symmetric square matrix, the error autocorrelation matrix. The diagonal elements of this matrix reflect the variance of the frequency offset error at each sampling time, while the off-diagonal elements reveal the statistical correlation of the error over time. In this way, random additive noise is significantly suppressed, while the systematic deviation characteristics with deterministic physical evolution are enhanced in the matrix.

[0068] The generated error autocorrelation matrix is ​​used as a dynamic observation noise covariance parameter and back-injected into the Kalman orbit tracking filter running within the high-orbit baseband processing unit. During each prediction-update cycle of the filter, this matrix replaces the fixed empirical noise parameter. In the update process, the processor calculates the observation matrix in real time, defined as the partial derivative matrix of the Doppler frequency offset with respect to the satellite's position and velocity state quantities. By mapping the Doppler error vector to the state space, the filter uses the Kalman gain to transform the frequency offset residual in the observation domain into an orbital element correction in the state domain. Subsequently, the filter executes the state update equation, using the currently observed frequency offset error to correct the state vector within the filter. Specifically, the state vector includes at least the three-dimensional position offset component and the three-dimensional velocity offset component of the target low-orbit satellite in the geocentric coordinate system. In this embodiment, the state vector also includes an augmented state quantity characterizing the local crystal oscillator frequency drift rate of the terminal, to achieve decoupled estimation of orbital offset and clock offset. In this process, the error autocorrelation matrix is ​​decomposed into eigenvalues ​​to extract the largest eigenvector that represents the six-root perturbation component of the space orbit. This quantifies the deviation of orbital parameters caused by the actual evolution of the satellite orbit (such as atmospheric drag and perturbation by the Earth's non-spherical gravitational force). In practice, the largest eigenvector is mapped to the six-root space of the orbit through a preset sensitivity matrix. For example, if the frequency offset error sequence is found to have a linear drift trend that increases constantly over time, the sensitivity matrix mainly maps it to the incremental correction of the local crystal oscillator frequency drift. If the frequency offset error is found to have an overall constant deviation or conforms to the Doppler rate of change (bell-shaped) characteristic, it is mainly mapped to the orbital position perturbation compensation caused by the semi-major axis. If the frequency offset error has a periodic oscillation characteristic, it is mainly mapped to the perturbation compensation of the eccentricity or perigee angle. By identifying the morphological characteristics of the error sequence in the time domain, the overall frequency offset residual is scientifically decomposed into independent correction values ​​corresponding to the semi-major axis perturbation, eccentricity perturbation, and local crystal oscillator frequency drift. Specifically, the preset sensitivity matrix is ​​obtained by performing a linear expansion on the satellite orbit prediction model, and its elements are defined as the partial derivatives of the theoretical Doppler frequency shift with respect to each space orbit parameter. During the mapping process, the processor performs a pseudo-inverse operation or least squares projection on the extracted maximum eigenvector (representing the main evolution trend of the error) and the sensitivity matrix. By identifying the morphological characteristics of the error sequence in the time domain (such as linear components, periodic components, and constant components), the overall frequency offset residual is decomposed into incremental correction values ​​corresponding to the semi-major axis perturbation, eccentricity perturbation, and local crystal oscillator frequency drift.

[0069] Finally, the extracted space orbit perturbation compensation is incrementally fed back into the locally running satellite orbit prediction model. This compensation is then used to perform weighted updates on key propagation model parameters in the ephemeris data, such as the semi-major axis, eccentricity, and perigee angle, generating a set of online-calibrated compensation weights. At the start of the next acquisition cycle, the orbit prediction model will recalculate the satellite position and velocity based on these updated parameters, thereby adaptively calibrating the pointing accuracy of the initial line-of-sight angle and achieving continuous autonomous optimization of the beam acquisition logic.

[0070] By constructing a dual-buffered data pipeline and a heterogeneous scheduling mechanism at the underlying hardware level, a seamless integration of highly complex compensation algorithms and high-speed radio frequency control flow was achieved. Utilizing a pipelined parallel architecture, the closed-loop latency from physical parameter sensing and model inference to beam control command issuance was significantly shortened, ensuring the real-time and continuous beam pointing of low-Earth orbit satellites during high-speed movement. This not only alleviated the computational load pressure caused by complex algorithms but also guaranteed signal coherence during beam switching.

[0071] Example 2: By locking the downlink broadcast channel of the high-orbit satellite through sliding related operations, the almanac parameters (ephemeris) of the target low-orbit satellite at the current moment and the global reference clock are demodulated and extracted. Through hardware latching logic, the counter value of the local controlled crystal oscillator is aligned with the high-orbit clock pulse. The local clock difference is eliminated by using a digital proportional-integral algorithm. At this time, the terminal obtains high-precision local reference time and preliminary satellite orbit prediction in advance without any low-orbit signal feedback.

[0072] As the vehicle travels across the bumpy sand, the inertial measurement unit senses the vehicle's pitch, roll, and heading changes in real time and aligns them with the antenna array coordinate system. Simultaneously, a network of temperature sensors installed at the bottom layer of the phased array captures the non-uniform temperature field generated by the combined heat from the desert's high temperature and the radio frequency amplifier, constructing a two-dimensional temperature gradient distribution table for the antenna array.

[0073] The main control processor cascades the extracted real-time pose angle (corresponding to spatial geometric changes), array surface temperature distribution (corresponding to RF thermal drift), and factory-preset RF fingerprint error (corresponding to process deviation) to reconstruct a multi-dimensional physical feature tensor, and sends it into the model input buffer.

[0074] The terminal invokes a nonlinear array compensation model integrated within the baseband chip. This model automatically identifies electromagnetic coupling between neighboring array elements and far-field interference across regions through a multi-scale dilated convolutional layer, and quantifies the contribution of each physical parameter to beam distortion through a spatial self-attention layer. The model inference outputs a phase predistortion matrix for the current complex operating conditions. Subsequently, based on the calculated theoretical satellite pointing, an ideal phase weight matrix is ​​generated and subjected to channel-level Hadamard product (element-level multiplication) with the model-output predistortion matrix. The fused beam control drive matrix is ​​then distributed to the digital phase shifters of each RF channel, enabling the antenna to synthesize a precisely pointing and stable physical receiving beam even under severe vehicle vibration and high-temperature drift.

[0075] After the physical receiving beam is accurately oriented to the line-of-sight direction of the low-Earth orbit satellite, the baseband processor uses theoretical Doppler frequency shift to perform reverse rotation compensation on the received signal. At the output of the sliding matched filter, the decision unit adopts a constant false alarm rate dynamic threshold, successfully extracting the correlation peak and locking the downlink synchronization signal even in an environment with drastic fluctuations in background noise.

[0076] Once the link is established, the terminal enters the closed-loop feedback phase. The processor extracts the actual Doppler observations generated by the phase-locked loop, subtracts them from the theoretical predictions to obtain the error sequence, and calculates the error autocorrelation matrix. This matrix reflects the drift residual between the almanac model and the actual operating state and is injected as a dynamic noise parameter into the Kalman orbit tracking filter. Through the state update equation, the system corrects the compensation weights of the local orbit prediction model in real time, thereby adaptively converging the beam pointing error as the vehicle continues to move.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive acquisition of low earth orbit satellite beams based on fusion dual architecture, characterized in that, Portable station terminals with built-in dual-mode baseband modules include: The almanac data and global reference clock of the target low-orbit satellite are extracted through the high-orbit communication link of the dual-system baseband module; the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna are collected. The local reference time is aligned using a global reference clock. Based on the local reference time, almanac data, and three-dimensional pose data, the initial line-of-sight angle and theoretical Doppler shift of the target low-orbit satellite relative to the phased array antenna are calculated, and the wide-angle off-axis angle is separated. The process involves: reading a pre-stored fingerprint error matrix; generating a phase predistortion matrix based on the wide-angle off-axis angle, operating temperature, and fingerprint error matrix; generating a theoretical pointing matrix based on the initial line-of-sight angle; and performing a Hadamard product operation between the theoretical pointing matrix and the phase predistortion matrix to generate a waveguide drive matrix. This includes: concatenating the wide-angle off-axis angle, operating temperature data, and fingerprint error matrix into a multi-dimensional physical feature vector; inputting the multi-dimensional physical feature vector into a pre-trained nonlinear array compensation model to output the phase predistortion matrix for the phased array antenna; calculating the ideal spatial path difference for each RF channel based on the physical element distribution topology of the phased array antenna, combined with the azimuth and elevation components of the initial line-of-sight angle, and generating the theoretical pointing matrix containing ideal phase weights; extracting the ideal phase elements of each channel from the theoretical pointing matrix and the corresponding conjugate compensation elements from the phase predistortion matrix, performing a Hadamard product operation on the corresponding elements, and generating a waveguide drive matrix using a channel-level mapping multiplication method. The wave control driving matrix includes comprehensive calibration weights; the nonlinear array compensation model includes a cascaded tensor construction layer, a multi-scale dilated convolutional layer, a spatial self-attention layer, and a feature tiling layer: the fingerprint error matrix is ​​read through the tensor construction layer and reconstructed into an initial feature tensor consistent with the two-dimensional physical arrangement dimension of the phased array antenna; the initial feature tensor is input to the multi-scale dilated convolutional layer, and feature extraction is performed in parallel using multiple sets of convolution kernels with different dilation rates to obtain the spatial electromagnetic coupling feature map between near-field adjacent array elements and far-field cross-region array elements inside the phased array antenna; the spatial electromagnetic coupling feature map is input to the spatial self-attention layer, the dot product correlation matrix between the feature vectors of each array element channel is calculated, and a spatial attention weight matrix is ​​generated based on the dot product correlation matrix; the spatial electromagnetic coupling feature map is multiplied element-wise with the spatial attention weight matrix, and a dimension reduction and flattening operation is performed through the feature tiling layer to output the physical coupling feature map; The beam control drive matrix is ​​input into the radio frequency front end of the phased array antenna to generate a physical receiving beam. Through the low-orbit communication link of the dual-mode baseband module, the downlink synchronization signal of the target low-orbit satellite is received using the physical receiving beam, and the downlink synchronization signal is analyzed. Extract the true Doppler frequency shift from the downlink synchronization signal, calculate the error autocorrelation matrix between the true Doppler frequency shift and the theoretical Doppler frequency shift, and inject the error autocorrelation matrix into the dual-mode baseband module to update the compensation weights of the ephemeris data.

2. The low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture according to claim 1, characterized in that, The built-in dual-mode baseband module includes a high-orbit baseband processing unit, a low-orbit baseband processing unit, and a data interaction bus; the high-orbit communication link is a narrowband signaling control link established based on the high-orbit baseband processing unit, used to maintain the underlying handshake with the auxiliary high-orbit satellite in a low signal-to-noise ratio environment; the low-orbit communication link is a broadband service data link established based on the low-orbit baseband processing unit. The data interaction bus is used to transmit the almanac data and global reference clock parsed by the high-orbit baseband processing unit to the low-orbit baseband processing unit in real time, as a priori compensation input for low-orbit signal acquisition.

3. The low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture according to claim 2, characterized in that, The process of acquiring data via the high-orbit communication link using a dual-mode baseband module includes: locking the downlink broadcast control channel of the auxiliary high-orbit satellite, demodulating the downlink broadcast control channel, parsing the orbital parameters of the target low-orbit satellite and the global reference clock signal used to calibrate the local oscillator, and using them as the almanac data; the acquisition of the three-dimensional pose data of the portable station terminal and the operating temperature data of the phased array antenna includes: measuring the pitch angle, roll angle, and yaw angle of the portable station terminal in real time through an inertial measurement unit rigidly connected to the phased array antenna to form the three-dimensional pose data; and acquiring the operating temperature data characterizing the thermal drift state of the radio frequency channel in real time through a temperature sensor network mounted on the bottom layer of the phased array antenna.

4. The low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture according to claim 1, characterized in that, The process of calculating the initial line-of-sight angle and theoretical Doppler frequency shift of the target low-orbit satellite relative to the phased array antenna, and separating the wide-angle off-axis angle, includes calibrating the local controlled oscillator inside the portable station terminal using the global reference clock signal, eliminating hardware clock bias, and outputting the local reference time. Based on the local reference time and the almanac data, the real-time spatial position vector and real-time velocity vector of the target low-orbit satellite in the preset geocentric coordinate system are calculated; the local carrier coordinate system of the phased array antenna is constructed according to the three-dimensional pose data, and the real-time spatial position vector is projected into the local carrier coordinate system using a coordinate transformation matrix to obtain the relative position vector, and the elevation and azimuth components of the relative position vector are extracted to form the initial line of sight; the real-time velocity vector is transformed into the local carrier coordinate system to obtain the relative velocity vector, the radial projection component of the relative velocity vector in the direction of the initial line of sight is calculated, and the theoretical Doppler frequency shift is generated based on the radial projection component and the communication carrier frequency; Calculate the true spatial angle between the initial line-of-sight angle and the physical normal vector of the phased array antenna. When it is determined that the true spatial angle is greater than a preset scanning gain attenuation threshold, the true spatial angle is extracted and separated as the wide-angle off-axis angle.

5. The low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture according to claim 1, characterized in that, The process of parsing the downlink synchronization signal includes: extracting the complex weight elements in the wave control drive matrix and mapping them down to each RF transceiver channel of the phased array antenna RF front end; configuring the feed phase and attenuation amplitude of the numerically controlled phase shifter and variable attenuator inside the corresponding RF transceiver channel using the complex weight elements, and synthesizing the physical receiving beam with a specific spatial pointing and sidelobe level through coherent superposition of space electromagnetic waves; capturing space electromagnetic signals using the physical receiving beam, performing low-noise amplification, down-conversion and analog-to-digital conversion operations on the space electromagnetic signals, and inputting the baseband digital signal into the low-Earth orbit communication link of the dual-system baseband module; performing sliding matched filtering with the baseband digital signal using the pre-stored local synchronization sequence in the low-Earth orbit communication link to extract the relevant peak energy; when the relevant peak energy is continuously detected to cross the preset acquisition decision threshold, confirming successful reception of the downlink synchronization signal of the target low-Earth orbit satellite, and parsing and extracting the physical layer frame number and timing synchronization identifier in the downlink synchronization signal.

6. The low-Earth orbit satellite beam adaptive acquisition method based on fused dual architecture according to claim 5, characterized in that, The process of updating the compensation weights of the almanac data includes: performing carrier frequency offset estimation and phase-locked loop residual extraction on the downlink synchronization signal to obtain frequency offset observations over multiple consecutive symbol periods to form the time series of the true Doppler frequency shift; performing a difference operation between the time series of the true Doppler frequency shift and the theoretical Doppler frequency shift to generate a time-domain Doppler error vector; performing an outer product operation between the time-domain Doppler error vector and its transpose to generate an error autocorrelation matrix; using the error autocorrelation matrix as the dynamic observation noise covariance, and injecting it inversely into the orbit tracking filter of the high-orbit baseband processing unit of the dual-system baseband module; using the state update equation of the orbit tracking filter to extract the space orbit perturbation compensation amount based on the eigenvalues ​​of the error autocorrelation matrix; using the space orbit perturbation compensation amount to update the local propagation model parameters of the almanac data to generate the updated compensation weights, so as to adaptively calibrate the initial line-of-sight angle in the next acquisition period.

Images