Low-orbit satellite communication anti-interference method based on RFSOC

Abstract

The application discloses a low-orbit satellite communication anti-interference method based on RFSOC, and comprises the following steps: collecting downlink signals by using a radio frequency direct sampling module of an RFSOC chip, estimating a Doppler rate and a total interference observation value in real time through multi-resolution time-frequency decomposition and a third-order extended state observer; feeding the Doppler rate to a Costas loop of a built-in model predictive controller to realize a prediction-correction type carrier synchronization, and dynamically configuring a wave trap filter set according to the total interference observation value to anchor and zero interference; inputting a high-dimensional state vector into a Monte Carlo tree search decision engine, and outputting a waveform strategy sequence through prospective game deduction with a virtual interference party; modulating a waveform parameter, triggering address remapping through a waveform index label, switching waveforms without interruption in one clock cycle, and generating an anti-interference uplink signal. The application realizes fast carrier locking and real-time interference suppression in a high dynamic environment, and improves the reliability and robustness of a low-orbit satellite communication terminal in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of satellite communication technology, and in particular to a method for anti-interference of low-Earth orbit satellite communication based on RFSOC. Background Technology

[0002] Low-Earth orbit (LEO) satellite communication, with its advantages of low latency, high bandwidth, global coverage, and low-cost deployment, has become a core technology for next-generation satellite internet, widely used in key areas such as military communications, emergency rescue, ocean shipping, and the Internet of Things (IoT). However, the high-speed motion of LEO satellites relative to ground terminals (typically 7.6 km / s) introduces significant Doppler frequency shifts and their rates of change. Simultaneously, the complex electromagnetic environment presents various types of interference, including narrowband targeting jamming, pulse jamming, broadband jamming, and increasingly prevalent intelligent mimicry jamming. These dynamic channel conditions, along with malicious interference, constitute the main technical obstacles to reliable LEO satellite communication terminal connections.

[0003] Traditional satellite communication terminals often employ an analog superheterodyne architecture combined with a digital baseband processing board. This approach has the following inherent drawbacks: First, the analog mixing, filtering, and intermediate frequency processing stages introduce significant noise and nonlinear distortion, and the multi-stage circuitry results in large size and high power consumption, making it difficult to meet the miniaturization requirements of the terminal. Second, the modem uses a pure feedback synchronization structure (such as a traditional Costas ring), which is prone to loss of lock or tracking delay in scenarios with Doppler change rates as high as several kHz / s. Third, anti-interference measures typically employ pre-set fixed filtering or frequency hopping modes, lacking real-time adaptive capabilities to unknown, time-varying interference; once the interference characteristics change, the communication link quality deteriorates sharply.

[0004] In recent years, with the maturity of radio frequency system-on-a-chip (RFSoC) technology, it has become possible to integrate radio frequency data converters (RFDCs), programmable logic (PLs), and high-performance processors (PSs) onto a single chip. However, existing RFSoC-based communication terminals are still mainly at the initial integration stage of "direct RF sampling replacing analog downconversion," failing to fully explore the potential of heterogeneous computing resources in real-time sensing, dynamic decision-making, and rapid reconfiguration. Specifically, the lack of a unified architecture capable of simultaneously handling high-dynamic Doppler, multiple interference types, and forward-looking waveform strategy planning results in terminals exhibiting technical bottlenecks such as "sensing lag, open-loop decision-making, and intermittent switching" when facing rapidly changing low-Earth orbit channels and intelligent interference.

[0005] Therefore, there is an urgent need for a low-Earth orbit satellite communication anti-interference method that can fully utilize the advantages of the RFSoC architecture to achieve high dynamic carrier synchronization, real-time interference suppression, and forward-looking waveform game theory. Summary of the Invention

[0006] The purpose of this invention is to provide a low-orbit satellite communication anti-interference method based on RFSOC, in order to solve at least some of the problems existing in the prior art.

[0007] Technical solution: A low-Earth orbit satellite communication anti-interference method based on RFSOC, comprising the following steps:

[0008] The downlink satellite communication radio frequency signal is acquired by the radio frequency direct acquisition module of the RFSOC chip, and multi-resolution time-frequency decomposition and extended state observation processing are performed to obtain a high-dimensional state vector containing real-time Doppler change rate and total interference observation value.

[0009] The real-time Doppler rate of change is fed forward to the predictive-correction Costas loop, and the NCO control word is predicted by the model predictive controller. At the same time, the notch filter bank is dynamically configured to set the interference anchor to zero based on the total interference observation value, so as to obtain the baseband demodulated data with interference suppression.

[0010] The high-dimensional state vector and the current waveform parameters are encoded into a state vector, which is then input into the Monte Carlo tree search decision engine to perform forward-looking multi-step game deduction and output a waveform strategy sequence.

[0011] The modulator on the RFSOC programmable logic side receives the waveform strategy sequence, preloads each waveform parameter in the sequence into different storage banks of the dual-port Block RAM, and inserts waveform index tags into the transmitted data frame.

[0012] When a waveform index tag is detected, the corresponding waveform parameters are switched through address remapping to generate a feedforward anti-interference modulation signal.

[0013] Beneficial effects: This invention achieves real-time, robust, and uninterrupted anti-interference transmission of low-orbit satellite communication in highly dynamic and complex electromagnetic environments through multi-resolution time-frequency sensing, predictive-correction carrier synchronization, and Monte Carlo tree search game decision-making. Attached Figure Description

[0014] Figure 1 This is a flowchart of the overall solution of the present invention.

[0015] Figure 2 This is a flowchart illustrating how the present invention obtains a high-dimensional state vector containing real-time Doppler change rate and total disturbance observations.

[0016] Figure 3 This is a flowchart illustrating the baseband demodulated data obtained by the present invention after interference suppression.

[0017] Figure 4 This is a flowchart of the output waveform strategy sequence of the present invention.

[0018] Figure 5This is a flowchart of the process for generating a feedforward anti-interference modulation signal according to the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0020] like Figure 1 As shown in the figure, this embodiment describes in detail the data processing flow of a low-Earth orbit satellite communication anti-interference method based on RFSOC, specifically including:

[0021] S1: The downlink satellite communication radio frequency signal is acquired using the RF direct acquisition module of the RFSOC chip, and multi-resolution time-frequency decomposition and extended state observation processing are performed to obtain a high-dimensional state vector containing the real-time Doppler change rate and total interference observation value.

[0022] In this embodiment, the radio frequency data converter (RFDC) module integrated within the RFSOC chip is first activated. This module directly performs bandpass sampling on the downlink radio frequency signal from the low-Earth orbit satellite at a sampling rate of several gigabits per second, outputting a parallel in-phase quadrature (IQ) digital baseband data stream. Specifically, the high-speed analog-to-digital converter inside the RFDC converts the analog radio frequency signal into a 12-bit precision IQ complex sequence, generating a set of I and Q components each clock cycle, which together constitute the original digital received signal.

[0023] Furthermore, the IQ data stream is simultaneously replicated and fed to three parallel pipeline channels with different processing delays and resource overheads to achieve multi-resolution time-frequency decomposition. The first channel is a high-speed path with a processing delay constrained to within 1 microsecond, used for real-time demodulation observations and subsequent extended state observations. The second channel is a medium-speed path with a delay ranging from 1 to 100 microseconds, performing a short-time Fourier transform (STFT) to convert the IQ data into time-frequency plot data, i.e., the energy distribution matrix of different frequency components at each time step. The third channel is a low-speed path with a delay greater than 100 microseconds, used to accumulate statistical characteristics, including but not limited to parameters describing the long-term distribution characteristics of the signal such as peak-to-average power ratio, cyclic stationary spectrum, and higher-order cumulants.

[0024] In a preferred embodiment, the real-time IQ data output from the high-speed path is used as input to a third-order extended state observer (ESO). This observer is essentially a nonlinear dynamic system that internally uses a nonlinear function fal(e,α,δ) to rapidly converge the error e between the current matched filter output and the observer's internal state prediction. Here, α is a nonlinear factor (within the range of 0 and 1), and δ is the width of the linear interval. At each time step, the extended state observer outputs three extended state components: the first component z1(k) represents the filtered useful signal; the second component z2(k) represents the real-time estimated Doppler rate of change (i.e., the first derivative of the carrier frequency with respect to time); and the third component z3(k) represents the total interference observation, which integrates narrowband interference, impulse interference, background noise, and unmodeled dynamic errors.

[0025] Based on a further improvement of this application, the three extended state components are aligned with the time-frequency graph data of the medium-speed path output and the statistical feature data of the low-speed path output according to strict timestamps, and then normalized and encoded to finally form a high-dimensional state vector S. t The dimension of this vector is typically in the tens to hundreds, and each element is dimensionless for subsequent processing by the decision engine.

[0026] S2: The real-time Doppler rate of change is fed forward to the prediction-correction Costas loop, and the NCO control word is predicted by the model prediction controller. At the same time, the notch filter bank is dynamically configured to set the interference anchor to zero based on the total interference observation value, so as to obtain the baseband demodulated data with interference suppression.

[0027] After obtaining the real-time Doppler rate of change z2(k), this scheme feeds it, along with the current phase error err(k) output from the conventional Costas ring phase detector, into a built-in Model Predictive Controller (MPC). This controller incorporates a simplified satellite orbital dynamics model, capable of predicting the Doppler frequency variation trend over several clock cycles based on the relative motion between the low-Earth orbit satellite and the ground terminal. Specifically, the MPC solves a small-scale quadratic programming problem with control magnitude and rate of change constraints in each clock cycle, with the optimization objective being to minimize the cumulative squared error between the predicted carrier phase and the received signal phase. The solution yields the optimal frequency control word sequence for the next L time steps, where the first element is the feedforward NCO drive signal for the current time step.

[0028] Preferably, the feedforward NCO drive signal and the output of a traditional second-order loop filter are dynamically weighted and fused. The fusion ratio is adaptively adjusted according to the current loop locking state: when the loop is not yet locked (e.g., during the initial acquisition phase), the weight of the feedforward component is set much larger than that of the feedback component to quickly pull in synchronization using prior information from orbital dynamics; when the loop is locked and the error is small, the weight of the feedback component is gradually increased to suppress observation noise. The fused signal yields the final NCO control word, which drives a local numerically controlled oscillator to generate a local synchronization carrier that is in phase and frequency with the received signal.

[0029] On the other hand, in the interference suppression process, the presence of strong interference is determined based on the total interference observation value z3(k) and its instantaneous frequency characteristics (obtained through zero-crossing detection or instantaneous frequency estimation algorithms). When |z3(k)| exceeds a preset dynamic threshold, it is determined that there is an interference component that needs to be actively filtered out at the current moment. Subsequently, the instantaneous frequency of the interference is calculated, and a set of notch center frequency control words is generated to quickly configure a notch filter bank consisting of multiple parallel dual second-order notch filters. This filter bank can dynamically zero out the interference frequency points while minimizing the impact on the useful signal frequency band. After notch processing, a clean IQ baseband signal with interference anchoring is output.

[0030] According to this embodiment, the clean IQ baseband signal is coherently demodulated with the local synchronization carrier: the in-phase branch and the quadrature branch are multiplied respectively, and then low-pass filtered to obtain the soft decision value of the baseband symbol. Then, the demodulated bit stream is obtained by symbol decision (e.g., for QPSK modulation, decision in four quadrants). At the same time, demodulation quality indicators, including error vector magnitude (EVM) and bit error rate estimate, are extracted. Thus, interference-suppressed baseband demodulated data is obtained.

[0031] S3: Encode the high-dimensional state vector and the current waveform parameters into a state vector, input it into the Monte Carlo tree search decision engine, perform forward-looking multi-step game deduction, and output the waveform strategy sequence.

[0032] This embodiment is used for interference-resistant waveform selection at the cognitive decision-making level. First, the high-dimensional state vector S from multiple consecutive time points (e.g., the past N time points) is... t-N+1 ,...,S t Aggregate coding is performed using the currently used waveform parameters (including modulation scheme, spreading factor, channel coding rate, pulse shaping roll-off factor, etc.) and the obtained demodulation quality indicators (such as EVM, bit error rate estimate). Aggregate coding employs concatenation and linear normalization to form a fixed-dimensional aggregated state S. t agg This aggregated state fully characterizes the current electromagnetic environment, channel dynamics, and the terminal's own configuration.

[0033] For example, the aggregation state S t agg The input is fed into a Monte Carlo Tree Search (MCTS) decision engine running on an ARM Cortex-A53 core on the RFSOC processing system (PS) side. Internally, this engine maintains a virtual interfering party model, which assumes the existence of an intelligent adversary aiming to minimize the user's communication throughput. The MCTS engine performs forward-looking multi-step game theory, as follows:

[0034] Selection: Starting from the root node (current state), select the optimal child node based on the number of visits and cumulative revenue of each candidate waveform action (i.e., the upper limit confidence interval criterion).

[0035] Expansion: After reaching the leaf node, expand according to the waveform action space (e.g., three modulation methods, BPSK, QPSK, and 8PSK, and two spreading factors, for a total of six candidate waveforms) and add new child nodes.

[0036] Simulation: Starting from the extended node, the simulation randomly simulates the adversarial process for the next K time steps (each time step corresponds to a transmission time slot, and the value of K is such that the total simulation time is 20-100 milliseconds). In the simulation, our side selects the waveform according to the current strategy, while the virtual jammer adopts a greedy strategy to select the jamming type that will reduce our side's gains the most (such as narrowband targeting jamming or pulse jamming).

[0037] Backpropagation: After the simulation ends, the revenue value is calculated based on the final bit error rate or throughput, and then propagated back along the path to update the access count and cumulative revenue of the waveform action on each node.

[0038] After multiple simulations (e.g., hundreds to thousands) within a single decision cycle, the probability distribution of each candidate waveform action accessed at the root node becomes the basis for decision-making. Random sampling is performed according to this probability distribution to sequentially generate waveform actions for the next K time steps, forming the waveform strategy sequence [W]. t+1 W t+2 ,...,W t+K ].

[0039] In one alternative implementation, the waveform policy sequence is bound and encapsulated with the future NCO control trajectory (i.e., the frequency control word sequence for the next L time steps) output internally by the Model Predictive Controller (MPC). The binding and encapsulation is performed by combining each waveform action with the average carrier frequency prediction value within the corresponding time window into a data structure, packaging it into a policy instruction package, and sending it from the PS side to the programmable logic (PL) side via shared memory or a dedicated hardware mailbox. This policy instruction package provides feedforward information to the modulation layer.

[0040] S4: The modulator on the RFSOC programmable logic side receives the waveform strategy sequence, preloads each waveform parameter in the sequence into different storage banks of the dual-port Block RAM, and inserts waveform index tags into the transmitted data frame.

[0041] When the modulator module on the RFSOC programmable logic side receives the policy instruction packet from the PS side, it first parses it to extract the waveform policy sequence and the detailed parameters corresponding to each waveform. These parameters include, but are not limited to: modulation map (constellation diagram coordinates), shaping filter coefficients (square root raised cosine roll-off coefficient), spreading code sequence, and transmit power amplitude factor.

[0042] According to a further improvement of this embodiment, the modulator internally provides a dual-port Block RAM resource pool, which is divided into multiple independent storage banks. Each bank stores a complete set of waveform parameters. The waveform parameters are preloaded into these banks in chronological order: the first waveform parameter is stored in Bank0, the second in Bank1, and so on, forming a circular buffer. The preloading process is controlled by the state machine on the PL side and will not interfere with the normal transmit data stream.

[0043] Simultaneously, waveform index tags are inserted into the uplink source-coded data stream to be transmitted. Specifically, a fixed-length field (e.g., 8 to 16 bits) is reserved after the preamble and before the payload in the physical layer frame structure. A timestamp synchronized by the model prediction controller triggers the generation of a tag value, which is equal to the bank number of the waveform parameter to be used. The tag value is written into this field to form a complete transmit data frame.

[0044] In a preferred embodiment, the generation of waveform index tags is synchronized with MPC to ensure that the waveform indicated by the tag is strictly aligned with the carrier trajectory predicted by the receiver. When the tag length is selected as 8 bits, up to 256 different waveform configurations can be supported, which is sufficient to cover the set of anti-interference waveforms commonly used in low-Earth orbit satellite communications.

[0045] S5: When a waveform index tag is detected, the corresponding waveform parameters are switched through address remapping to generate a feedforward anti-interference modulation signal.

[0046] This embodiment utilizes pre-loaded waveform parameters to achieve uninterrupted waveform switching and generate uplink RF signals. First, the transmitted data frame is sent to the waveform switching control unit. This unit continuously monitors the field content following the preamble as the data frame passes through. Once a valid waveform index tag value is detected, the switching operation is immediately triggered.

[0047] The core of the switching operation is the address remapping mechanism. Specifically, a pre-configured read-only lookup table (LUT) takes the detected tag value as input and outputs the high-order bits of the corresponding Block RAM read address. Since different banks in the Block RAM are distinguished by their high-order bits in the address space, changing these high-order bits achieves bank switching. The entire switching process is completed within one clock cycle, without interrupting the transmit data pipeline or reloading waveform parameters. After the switching is complete, the waveform switching control unit outputs a switching completion flag, which can be used to trigger the prefetching of waveform parameters for the next bank, forming a pipelined operation.

[0048] After obtaining the currently valid waveform parameters, the modulator digitally modulates the baseband data from the source code. Taking QPSK modulation as an example: the input bitstream is converted from serial to parallel to obtain two bit paths, which are mapped to sign values ​​(e.g., +1 or -1) for the I and Q paths, respectively. These are then convolved with the shaping filter coefficients to achieve pulse shaping and reduce inter-symbol interference. The shaped digital baseband signal is multiplied by the carrier signal (generated by the NCO) through a multiplier to obtain the digital intermediate frequency modulated signal.

[0049] Subsequently, the digital intermediate frequency (IF) modulated signal is fed into the DAC module integrated into the RFSOC chip. The DAC module first performs digital predistortion compensation, and then, through digital-to-analog conversion, low-pass filtering, and amplification, restores the digital signal to an analog IF signal. This analog IF signal then passes sequentially through an external IF filter (used to filter out image frequencies), an RF mixer (shifting the IF to the specified uplink RF frequency), an RF filter (suppressing out-of-band spurious signals), and a power amplifier, finally being radiated out by the antenna to form an uplink anti-interference communication RF signal that can be received by low-Earth orbit satellites.

[0050] Furthermore, this scheme includes a closed-loop feedback mechanism: it collects bit error rate statistics from the baseband demodulated data stream and extracts the trend of interference power changes from the total interference observations output by the extended state observer. These actual communication quality data are then fed back to the Monte Carlo tree search decision engine on the PS side to update the parameters of the virtual interference model and the payoff function online (e.g., adjusting the bit error rate-to-payoff mapping curve), thereby achieving continuous optimization of the decision model and enabling the terminal to adapt to unknown or evolving electromagnetic interference environments.

[0051] According to another aspect of this application, such as Figure 2 As shown, a high-dimensional state vector containing the real-time Doppler rate of change and the total disturbance observations is obtained, specifically including:

[0052] S101: The IQ data stream output by the RFDC is sent to three parallel processing channels: the first channel is a high-speed path for real-time demodulation and observation; the second channel is a medium-speed path for short-time Fourier transform; and the third channel is a low-speed path for statistical feature accumulation.

[0053] In this embodiment, the parallel IQ data stream output from the radio frequency data converter (RFDC) inside the RFSOC chip is first copied into three identical copies, which are then sent to three independent processing channels. The first channel is configured as a high-speed path, and its hardware implementation uses pure combinational logic and a minimal pipeline, without introducing additional buffers or complex calculations, thereby strictly controlling the processing latency to within 1 microsecond. The first channel is used for real-time demodulation observation, that is, directly passing the IQ data to the subsequent extended state observer without any downsampling or filtering processing, to ensure the original response to rapidly changing channel dynamics.

[0054] Furthermore, the second channel is a medium-speed path with a delay between 1 microsecond and 100 microseconds. This path integrates a Short-Time Fourier Transform (STFT) processing unit, employing a sliding window approach to frame, window (e.g., Hanning window), and perform Fast Fourier Transform (FFT) on the IQ data stream. Each transform outputs a complex spectral vector, the square of which represents the time-frequency plot data for that time period. The time-frequency plot is stored in matrix form, with rows representing frequency bins and columns representing time frames, providing a visual representation of the energy distribution of signals and interference on the time-frequency plane.

[0055] Furthermore, the third channel is a low-speed path with a delay greater than 100 microseconds. This path does not aim for real-time processing but rather accumulates statistical features over a longer period. Specifically, it includes a statistical feature extraction engine that calculates the peak-to-average power ratio (PAPR), cyclic stationary spectrum (obtained through cyclic autocorrelation processing of the signal), and higher-order cumulants (such as skewness and kurtosis) frame by frame. These statistical features do not depend on instantaneous values ​​but reflect long-term patterns in the interference environment, such as the presence of periodic impulse interference or non-stationary noise.

[0056] In a preferred embodiment, the final outputs of the three channels all have a unified timestamp label for alignment and fusion in subsequent steps.

[0057] S102: Input the real-time IQ data output from the high-speed path to the third-order extended state observer. The nonlinear function is used to quickly converge the error between the current matched filter output and the observer's prediction value, and output three extended states: the filtered useful signal, the real-time Doppler rate of change estimate, and the total interference observation value.

[0058] According to a further improvement in this application, the real-time IQ data output from the high-speed path is used as a dynamic input and fed into a third-order extended state observer (ESO). The observer is designed based on active disturbance rejection control theory, with its core state equation in discrete-time form. The observer maintains three state variables, corresponding to the filtered signal, the Doppler rate of change, and the total disturbance, respectively. At each sampling time, the observer first uses the state from the previous time step to predict the output at the current time step, and then subtracts the predicted value from the actual input IQ data to obtain the error signal e(k).

[0059] Preferably, the error e(k) is amplified or compressed using a nonlinear function fal(e,α,δ). This function is defined as: fal(e,α,δ)={|e| α ·sgn(e),∣e∣>δ;e / δ 1-α , |e|≤};where α is the nonlinearity factor (typically 0.5) and δ is the width of the linear interval (typically 0.01). When the error is large, the function behaves as a power function, achieving fast convergence; when the error is small, it degenerates into a linear proportional element to avoid high-frequency chatter. The error processed by the fal function is multiplied by three gain coefficients to update the three state variables of the observer.

[0060] Specifically, the observer outputs three extended states: the first state z1(k) represents the filtered useful signal, which has removed most of the high-frequency noise and interference; the second state z2(k) represents the real-time Doppler rate of change, i.e., the first derivative of the carrier frequency with respect to time, measured in Hertz per second; and the third state z3(k) represents the total interference observation, a lumped term that includes the combined effects of narrowband interference, impulse interference, unmodeled dynamics, and model errors. These three states are updated once per sampling period and passed to subsequent steps.

[0061] S103: Align and normalize the three expansion states with the time-frequency graph data of the medium-speed path output and the statistical feature data of the low-speed path output according to the timestamp, and encode them into a high-dimensional state vector.

[0062] After obtaining the three expanded states, they need to be fused with the time-frequency plot data of the medium-speed path and the statistical characteristic data of the low-speed path. Since the processing delays of the three channels are different, direct splicing would lead to time mismatch. Therefore, a timestamp alignment operation is performed first. Each data point is marked with a count value based on a high-precision timer within the RFSOC when it is generated. The alignment module maintains a small FIFO buffer to temporarily store data from channels with shorter delays, waiting for data from channels with longer delays to arrive, ensuring that data within the same time window are retrieved together.

[0063] Next, normalization is performed. For the extended state z1(k) (the useful signal), its amplitude range depends on the received signal power; it is divided by a long-term average amplitude to obtain a dimensionless relative value. For the Doppler rate of change z2(k), its theoretical range is known (e.g., the maximum Doppler rate of change for low-Earth orbit satellites is approximately ±10 kHz / s), and a linear mapping is used to the [-1,1] interval. For the total interference observation z3(k), normalization is also performed based on the historical maximum value. The time-frequency plot data is a two-dimensional matrix, and the value range of each element is related to the ADC quantization bit depth; it is uniformly divided by the full-scale value. Statistical characteristics such as peak-to-average power ratio (PAPR) are already dimensionless; only checks for out-of-range values ​​are needed.

[0064] According to this embodiment, all normalized data are concatenated into a one-dimensional vector according to a predefined dimensional order, which is the high-dimensional state vector S. t The vector may have tens of dimensions (e.g., 3 dimensions for the expanded state, 16 dimensions for the time-frequency plot after principal component compression, and 10 dimensions for statistical features, totaling 29 dimensions). This vector serves as the input feature for the subsequent Monte Carlo tree search decision engine, providing a complete description of the current electromagnetic environment and channel state.

[0065] According to another aspect of this application, such as Figure 3 As shown, the baseband demodulated data with interference suppression is obtained, specifically including:

[0066] S201: Input the real-time Doppler change rate and the current phase error to the model predictive controller to predict the optimal frequency control word and obtain the feedforward NCO drive signal.

[0067] This step details the feedforward prediction mechanism of the Model Predictive Controller (MPC). The real-time Doppler rate of change z2(k) output from the extended state observer and the current phase error err(k) from the conventional Costas ring phase detector are simultaneously input to the MPC module. Internally, the MPC encapsulates a simplified satellite orbital dynamics model that describes the time-varying relative radial velocity between the low-Earth orbit satellite and the ground terminal. Specifically, it is represented by a second-order differential equation: Doppler acceleration is related to orbital parameters. In the discrete domain, this model is expressed as a state transition matrix.

[0068] Preferably, the MPC performs the following operations in each clock cycle: First, using the current z2(k) as the initial state, it substitutes the state transition matrix to predict the Doppler frequency values ​​for the next L times (e.g., L=5 to 10). Then, it combines the predicted frequency values ​​with the current phase error err(k) to construct a quadratic cost function, which includes the sum of the squares of the carrier phase differences at future times and a penalty term for the control word change. By solving a constrained quadratic programming problem (the control word change rate cannot exceed the maximum frequency step of the DAC), the optimal control word sequence {u} that minimizes the cost function is obtained. k ,u k+1 ,...,u k+L-1}

[0069] For example, the first element u of the sequence k This is the feedforward NCO drive signal at the current moment. This signal does not rely on traditional loop filter feedback, but is based on forward calculations using a dynamic model, enabling it to compensate for frequency offsets caused by Doppler variations in advance. Compared to a pure feedback Costas loop, this feedforward structure can reduce carrier tracking error by an order of magnitude in highly dynamic environments (such as during low-Earth orbit satellite overhead).

[0070] S202: The feedforward NCO drive signal and the loop filter output are dynamically weighted and fused to obtain the NCO control word, which drives the local carrier generation to achieve predictive-corrective phase update.

[0071] After obtaining the feedforward NCO drive signal, this scheme performs weighted fusion with the output of a traditional Costas second-order loop filter. The loop filter performs low-pass filtering and integration on the phase error of the phase detector output to generate a feedback frequency control signal. The fusion of the two signals adopts a dynamic proportional weighting method, that is, the final NCO control word ctrl(k) satisfies: ctrl(k) = α(k)·ctrl ff (k)+(1-α(k))·ctrl fb (k); where ctrl ff (k) represents the feedforward component, ctrl fb (k) is the feedback component, and α(k) is the adaptive fusion coefficient, with a value range of [0,1].

[0072] According to a further improvement in this embodiment, the fusion coefficient α(k) is adaptively adjusted based on the current loop lock state. The lock state can be estimated using the error variance output by the phase detector: when the variance is large (loop not locked, such as during the initial acquisition phase or after being lost due to strong interference), α(k) is set to close to 1, so that the NCO mainly follows the feedforward prediction value and quickly pulls into synchronization using prior information on orbital dynamics. When the error variance decreases (loop locked), α(k) is gradually decreased, and the weight of the feedback component is increased to utilize the noise suppression characteristics of the loop filter to obtain smoother tracking.

[0073] The final generated NCO control word is written to the frequency control register of the numerically controlled oscillator. The NCO's internal accumulator uses this control word as a step to generate a local in-phase carrier cos(2πf). c t) and orthogonal carrier sin(2πf c The t) are sent to the two multiplication branches of the demodulator, respectively, realizing a phase update mechanism that combines prediction and correction.

[0074] S203: Based on the total interference observation value and its instantaneous frequency characteristics, calculate the instantaneous frequency of the interference and generate the notch center frequency control word to obtain the clean IQ baseband signal after interference anchoring.

[0075] This step is used for active interference anchoring. The total interference observation z3(k) output by the extended state observer is a time-varying scalar, but it does not contain frequency information itself. To obtain the instantaneous frequency of the interference, frequency features need to be extracted from the IQ data. One approach is to combine z3(k) with the original IQ data and process it through an instantaneous frequency estimator. This estimator first differentiates the IQ data and then calculates the phase difference to obtain the instantaneous angular frequency.

[0076] In a preferred embodiment, |z3(k)| only exceeds a dynamic threshold Th dyn Interference frequency estimation is triggered only when the background noise power is exceeded. The dynamic threshold is adaptively adjusted based on the background noise power, for example, set to 3 times the long-term average signal amplitude. Exceeding the threshold indicates the presence of significant external interference (such as narrowband targeted interference or pulse interference). At this time, the instantaneous frequency estimation algorithm outputs a frequency value f. jam (k).

[0077] The frequency value is then converted into the center frequency control word for the notch filter. The notch filter bank consists of multiple parallel biquad notch filters, each with independently configurable center frequency, quality factor, and notch depth. The control word is mapped to the filter coefficient update values ​​via a lookup table. Since the update process involves only a few multiply-accumulate operations, the new coefficients can be loaded within a few clock cycles. After configuration, the IQ data stream passes through the notch filter bank, attenuating the energy at the corresponding interference frequency by more than 30dB, while the useful communication signal remains largely unaffected. The filtered output is the clean IQ baseband signal after interference anchoring.

[0078] S204: Perform coherent demodulation and symbol decision on the clean IQ baseband signal and the local synchronization carrier to obtain interference-suppressed baseband demodulated data, and extract demodulation quality indicators.

[0079] The clean IQ baseband signal, after notch filtering, has significantly reduced interference components. Next, coherent demodulation is performed: the clean IQ signal is compared with the in-phase component cos(2πf) of the local synchronization carrier. c t) and orthogonal components sin(2πf c The products are multiplied by t, and the product is filtered out by a low-pass filter (e.g., a moving average or FIR filter) to remove the second harmonic component, yielding the baseband in-phase symbol I. b (k) and the orthogonal symbol Q b (k).

[0080] For typical phase shift keying modulation (such as BPSK, QPSK, 8PSK), the symbol decision device is based on I... b (k) and Q b (k) The position on the constellation diagram determines the corresponding bit combination. Taking QPSK as an example, the decision region has four quadrants: if I b >0 and Q b >0, the decision is bit "00"; I b <0 and Q b If the value is >0, the decision is "01"; and so on. The bit stream output by the decision is the baseband demodulated data after interference suppression.

[0081] Simultaneously, two important quality metrics are extracted during demodulation: Error Vector Magnitude (EVM) and Bit Error Rate (BER) estimate. EVM is obtained by calculating the root mean square of the Euclidean distance between the actual symbol points and the ideal constellation points, reflecting modulation accuracy and residual interference levels. BER estimate utilizes the known pilot sequence in the frame header to statistically determine the percentage of bits with incorrect decisions. These two metrics, encapsulated together with the demodulated data, are output to subsequent steps for closed-loop optimization of the cognitive decision engine.

[0082] According to another aspect of this application, such as Figure 4 As shown, the specific steps of the output waveform strategy sequence include:

[0083] S301: Based on the high-dimensional state vector, demodulation quality index, and the waveform parameters currently in use, perform state aggregation encoding to obtain the aggregated state.

[0084] This step describes how to fuse information from multiple sources into a single aggregated state. First, three types of input data are collected: the first type is a high-dimensional state vector S spanning multiple consecutive time points (e.g., the past N=5 time points). t-4 ,...,S t The first category consists of two scalars: the first is the demodulation quality index at the current moment, which includes EVM and bit error rate estimation; the second category consists of the waveform parameters currently in use, which include the modulation scheme coding (e.g., 0 represents BPSK, 1 represents QPSK, and 2 represents 8PSK), the spreading factor (integer), and the coding rate (floating-point number).

[0085] According to this embodiment, the aggregation encoding consists of three steps: concatenation, normalization, and dimensionality reduction. First, all data is concatenated into an original long vector of length 5×D+5. Since D may be large (e.g., 29 dimensions), the length of the concatenated vector is approximately 150 dimensions, which would incur a computational burden if used directly. Therefore, the second step involves z-score normalization (subtracting the mean and dividing by the standard deviation) for each dimension to ensure that all components have the same scale. The third step uses a pre-trained autoencoder or principal component analysis (PCA) projection matrix to reduce the dimensionality of the normalized vector to a fixed low-dimensional space (e.g., 16 or 32 dimensions), obtaining the aggregation state S. tagg The aggregated state retains the main variance information of the original data, while reducing the input dimensionality of the subsequent MCTS decision engine.

[0086] S302: Input the aggregated state into the Monte Carlo tree search decision engine. The engine maintains a virtual interference model and performs game theory simulation: starting from the current state, it simulates the confrontation process between our candidate waveform strategy and the virtual interference optimal interference strategy in the next K time steps. Using the bit error rate or throughput as the payoff function, it backpropagates to update the access count and cumulative payoff of each waveform action, and obtains the access probability distribution of each candidate waveform action.

[0087] The aggregation state S taggThe input is fed into a Monte Carlo Tree Search (MCTS) decision engine running on the ARM core of the RFSOC processing system. This engine internally maintains a virtual jammer model, which is a conditional probability distribution. Given our waveform action, it outputs the most likely jamming type (e.g., narrowband jamming, pulse jamming, frequency sweep jamming, etc.) that the jammer will employ. The virtual jammer's strategy is set to a "greedy strategy that minimizes our gains," meaning it selects the jamming method that results in the worst expected communication quality each time.

[0088] MCTS performs multiple rounds of simulation, each consisting of four phases. First, the selection phase: starting from the root node (corresponding to the current state), child nodes are recursively selected until a leaf node is reached. The selection is based on the Upper Limit Confidence Interval (UCT) criterion, balancing exploration and exploitation. Second, the expansion phase: at the leaf node, new child nodes are added based on the candidate waveform action space (e.g., 6 waveform combinations). Third, the simulation phase: starting from the new node, the adversarial process for the next K time steps is randomly simulated. In each simulation step, the player selects waveforms according to the current strategy, while the virtual jammer selects jamming according to a greedy strategy. Then, the bit error rate (BER) or throughput for that step is calculated based on a pre-trained reward function. The reward function is designed to be the negative logarithm of the BER (so that lower BER results in higher rewards) or a linear function of the throughput. Fourth, the backpropagation phase: the accumulated rewards obtained from the simulation are propagated back along the path layer by layer, updating the access count and accumulated rewards for each waveform action at each node.

[0089] Within a single decision cycle (e.g., 10 milliseconds), the engine repeats the above simulation process thousands of times. Ultimately, the probability distribution of the access frequency of each candidate waveform action under the root node represents its probability distribution. This distribution reflects the relative merits of each waveform action after adversarial evaluation.

[0090] S303: Randomly sample according to the access probability distribution, and generate waveform actions for the next K time steps in sequence to obtain the waveform strategy sequence.

[0091] After obtaining the access probability distribution for each candidate waveform action, this step transforms the probability distribution into a specific waveform action sequence. Unlike traditional deterministic decision-making (always choosing the action with the highest probability), this scheme employs a random sampling strategy to increase unpredictability, thereby countering intelligent interference.

[0092] Specifically, first, a cumulative distribution function (CDF) is generated based on the probability distribution. Then, for each of the next K time steps (k=1..K), a random number uniformly distributed between 0 and 1 is generated, and the corresponding waveform action is determined by looking up the CDF. The value of K ensures that the total simulation time is between 20 and 100 milliseconds, with the specific value depending on the overhead velocity of the low-Earth orbit satellite and the time-varying characteristics of the interference environment. For example, if K=10, with each time step being 5 milliseconds, then the total look-ahead window is 50 milliseconds.

[0093] K waveform actions are sampled sequentially and arranged in chronological order to form a waveform strategy sequence [W]. t+1 W t+2 ,...,W t+K The characteristic of this sequence is its non-determinism: with the same input state, two decisions may produce different output sequences, which increases the difficulty for the perturber to learn.

[0094] S304: Bind and encapsulate the waveform strategy sequence with the future NCO control trajectory predicted by the model predictive controller to obtain the strategy instruction packet sent to the modulation layer.

[0095] To enable the modulation layer to prepare waveform parameters in advance and achieve feedforward carrier synchronization, the waveform strategy sequence needs to be bound and encapsulated with the future NCO control trajectory output by the Model Predictive Controller (MPC). The MPC has already calculated the optimal frequency control word sequence {u} for the next L time steps. k+1 ,...,u k+L This sequence describes the predicted trajectory of future carrier frequency changes.

[0096] Since the time window of the waveform policy sequence (K time steps, each potentially corresponding to multiple sampling periods) and the MPC prediction window (L sampling periods) may not be consistent in time scale, interpolation or alignment is required. One implementation method is to use the average frequency prediction value of multiple sampling periods within each time step as the carrier reference frequency for that time step. Then, the waveform action and the carrier reference frequency for that time step are combined into an atomic unit. All K atomic units are sequentially packaged into a data packet, i.e., the policy instruction packet.

[0097] The policy instruction packet structure includes: a header (identifying the instruction type), a sequence length K, and K entries, each containing waveform action code (e.g., 3 bits), waveform parameter index (e.g., 8 bits), and carrier frequency offset prediction value (16 bits). This instruction packet is transmitted from the PS side to the modulator module on the PL side via a high-speed bus (e.g., AXI4-Stream) inside the RFSOC.

[0098] According to another aspect of this application, such as Figure 5As shown, the specific steps for generating a feedforward anti-interference modulation signal include:

[0099] S501: Parse the strategy instruction packet to obtain the waveform strategy sequence and the corresponding waveform parameters, and preload the waveform parameters into different memory banks of the dual-port Block RAM in chronological order to obtain the memory bank resource pool of loaded waveform parameters.

[0100] When the modulator on the PL side receives the policy instruction packet, it first parses it. The parsing module verifies the integrity of the packet and extracts the sequence length K and the waveform parameters of each entry. The waveform parameters include: modulation type (BPSK / QPSK / 8PSK), shaping filter coefficient array (e.g., 64 floating-point coefficients), spreading code sequence (e.g., 128 bits), and transmit power scaling factor.

[0101] In a preferred embodiment, the modulator internally includes a dual-port Block RAM, which is divided into M independent banks, where M is at least equal to K (typically M=16 or 32). Each bank has independent write and read address ports, supporting simultaneous writing and reading. Waveform parameters are preloaded sequentially: the waveform parameters for the first future time step are written to Bank0, the second to Bank1, and so on, up to Bank(K-1). The preloading process is controlled by a finite state machine on the PL side, using the write port, and does not interfere with the ongoing transmission operation (read port). After all writing is complete, the data structure within each bank remains consistent: the first few addresses store filter coefficients, followed by the spreading code, and finally the control parameters. This forms a pool of memory resources with loaded waveform parameters, awaiting subsequent index tag trigger switching.

[0102] S502: Insert the waveform index tag into the uplink source-coded data stream to be transmitted to obtain the transmit data frame carrying the waveform index tag.

[0103] Along the transmission path, the uplink source-coded data stream to be transmitted (which has already undergone source compression and channel coding) is organized in frames. Each frame consists of a preamble, a frame header, a payload, and a check field. According to this scheme, a dedicated waveform index field with a length of 8 to 16 bits is designed in the frame header.

[0104] The generation of waveform index tags is synchronized with system time and the prediction window of MPC. Specifically, at the beginning of each transmit slot, the modulator obtains the Bank number corresponding to the waveform action to be used from the policy instruction packet and writes this number directly as the tag value into the waveform index field of the frame header. The position of this field is fixed after the preamble and before the payload to ensure that the receiver can parse the waveform index before demodulating the valid data. The complete data frame after inserting the tag is the transmit data frame carrying the waveform index tag.

[0105] For example, to enhance robustness, waveform index tags can be sent twice (with adjacent fields being identical) or redundantly in multiple frames so that the receiver can still correctly identify them under interference conditions.

[0106] S503: Sends the transmitted data frame to the waveform switching control unit. When the waveform index tag is detected, the waveform index tag value is mapped to the high bit of the read address of the Block RAM through a pre-configured lookup table. The memory switching is completed within one clock cycle to obtain the current valid waveform parameters and output the switching completion flag.

[0107] The transmitted data frames pass bit-by-bit through the waveform switching control unit in the pipeline. This unit contains a tag detector that continuously monitors the data stream for the field following a predefined preamble end marker. Once a valid waveform index tag value is detected (e.g., a tag value ranging from 0 to M-1), the switching logic is immediately triggered.

[0108] The core of the switching logic is a pre-configured read-only lookup table (LUT). The LUT takes a tag value as input and outputs the high-order bits of the corresponding Block RAM read address. Since different banks of Block RAM are distinguished by their high-order addresses in the address space (e.g., if each bank is 1024 bytes in size, the high 4 bits determine the bank number), changing the high-order bits switches banks. The LUT's output is directly connected to the high-order port of the Block RAM read address, while the low-order address is provided by the byte count of the transmitted data stream. Because both the LUT and the address lines are combinational logic, the switching operation is completed within one clock cycle, without causing pipeline stalls.

[0109] After the switching is complete, the waveform switching control unit outputs a single-clock-cycle switching completion flag pulse. This pulse can be used to trigger two actions: first, to lock the currently valid waveform parameters for subsequent modulation; and second, to prefetch data from the next bank (if any), forming a pipelined cascade.

[0110] S504: Use the current effective waveform parameters to perform digital modulation and shaping filtering on the baseband data to be modulated, and obtain a digital intermediate frequency modulated signal.

[0111] After obtaining the current valid waveform parameters, the modulator begins processing the baseband data. First, based on the modulation type in the waveform parameters, the input bitstream is mapped into a sequence of complex symbols. For QPSK, two bits are grouped together and mapped to I / Q symbols, such as "00" -> (1,1), "01" -> (-1,1), etc. For 8PSK, three bits are mapped to one of eight phase points. The mapped symbol sequence is then XORed or multiplied with the spreading code (if any) in the waveform parameters for spread spectrum generation.

[0112] Shaping filtering is then performed. The shaping filter uses a square root raised cosine (RRC) filter, whose coefficients are pre-stored in the current bank of the Block RAM. The symbol sequence is input to the filter at the symbol rate, and the filter outputs the interpolated waveform at a sampling rate several times the symbol rate (e.g., 4x oversampling). Through convolution operations, a baseband waveform with limited bandwidth and minimal intersymbol interference is obtained.

[0113] Finally, digital up-conversion is achieved by multiplying the baseband waveform with the carrier signal. The carrier signal is generated by the NCO, whose frequency control word is provided by a predictive-corrective loop, but also receives feedforward frequency offset compensation from the policy instruction packet. The product is the digital intermediate frequency modulated signal, which is output at the DAC interface rate within the RFSOC.

[0114] S505: The digital intermediate frequency modulation signal is sequentially sent to the DAC module built into the RFSOC for digital-to-analog conversion, and then processed by intermediate frequency filtering, radio frequency mixing, radio frequency filtering and power amplifier to obtain an anti-interference uplink communication radio frequency signal radiated out by the antenna.

[0115] The digital intermediate frequency (IF) modulated signal undergoes digital-to-analog conversion (DAC) via the DAC submodule within the RFDC module of the RFSOC. The DAC module first performs digital predistortion (DPD) compensation on the digital signal to overcome power amplifier nonlinearity, and then converts it into an analog IF voltage signal at a sampling rate of several GHz. This analog signal is then reconstructed and filtered by an on-chip programmable low-pass filter (LPF) to remove the sampling image frequencies.

[0116] Next, the signal is led out from the output pin of the RFSOC chip to the off-chip RF front-end circuit. First, it enters an intermediate frequency filter (typically a surface acoustic wave filter or dielectric filter) to further filter out out-of-band spurious signals. Then, it is fed into an RF mixer to mix with a high-frequency carrier from the local oscillator (e.g., an uplink frequency of 14 GHz), shifting the signal spectrum to the specified uplink RF band. The mixer output passes through an RF bandpass filter to suppress image frequencies and local oscillator leakage generated during mixing. Finally, the signal is amplified to the required transmit power by a power amplifier and radiated into free space via an antenna, forming an uplink communication signal that can be received by low-Earth orbit satellites.

[0117] Because the entire link employs predictive-correction carrier synchronization, dynamic notch filtering for interference suppression, and forward-looking waveform strategy sequences, the final output uplink RF signal has the ability to maintain reliable communication in complex electromagnetic environments.

[0118] This invention achieves real-time joint estimation of high-dynamic Doppler change rate and unknown interference type through the coordinated operation of multi-resolution time-frequency decomposition and a third-order extended state observer, solving the problem of easy loss of lock-up in traditional schemes when low-Earth orbit satellites pass overhead at high speeds. Furthermore, a model predictive controller is introduced into the Costas loop, replacing pure feedback tracking with a feedforward-feedback dynamic weighted fusion method, improving carrier synchronization acquisition speed and tracking accuracy in complex channels. Additionally, a Monte Carlo tree search-based decision engine incorporates a virtual interference party model, outputting a nondeterministic waveform strategy sequence through forward-looking multi-step game deduction, enabling the terminal to proactively combat intelligent interference rather than passively responding. Moreover, waveform parameters are pre-loaded into the Block. The RAM, combined with tag-triggered address remapping, enables uninterrupted waveform switching within a single clock cycle, eliminating link interruptions and data loss caused by waveform switching in traditional solutions. Finally, the overall architecture is deeply adapted to the heterogeneous resources of RFSoC, with direct RF sampling, real-time processing of programmable logic, and ARM core cognitive decision-making forming a tightly coupled pipeline, reducing system latency, size, and power consumption. This provides a highly reliable, real-time, and adaptive anti-interference solution for low-Earth orbit satellite communication terminals in military and civilian fields.

[0119] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for anti-interference of low-Earth orbit satellite communication based on RFSOC, characterized in that, Includes the following steps: The downlink satellite communication radio frequency signal is acquired by the radio frequency direct acquisition module of the RFSOC chip, and multi-resolution time-frequency decomposition and extended state observation processing are performed to obtain a high-dimensional state vector containing real-time Doppler change rate and total interference observation value. The real-time Doppler rate of change is fed forward to the predictive-correction Costas loop, and the NCO control word is predicted by the model predictive controller. At the same time, the notch filter bank is dynamically configured to set the interference anchor to zero based on the total interference observation value, so as to obtain the baseband demodulated data with interference suppression. The high-dimensional state vector and the current waveform parameters are encoded into a state vector, which is then input into the Monte Carlo tree search decision engine to perform forward-looking multi-step game deduction and output a waveform strategy sequence. The modulator on the RFSOC programmable logic side receives the waveform strategy sequence, preloads each waveform parameter in the sequence into different storage banks of the dual-port Block RAM, and inserts waveform index tags into the transmitted data frame. When a waveform index tag is detected, the corresponding waveform parameters are switched through address remapping to generate a feedforward anti-interference modulation signal.

2. The method according to claim 1, characterized in that, We obtain a high-dimensional state vector containing the real-time Doppler rate of change and the total disturbance observations, specifically including: The IQ data stream output from the RFDC is fed into three parallel processing channels: the first channel is a high-speed path for real-time demodulation and observation; the second channel is a medium-speed path for short-time Fourier transform; and the third channel is a low-speed path for statistical feature accumulation. The real-time IQ data output from the high-speed path is input into a third-order extended state observer. The error between the current matched filter output and the observer's prediction is quickly converged through a nonlinear function, and three extended states are output: the filtered useful signal, the real-time Doppler rate of change estimate, and the total interference observation. The three expansion states are aligned with the time-frequency graph data of the medium-speed path output and the statistical feature data of the low-speed path output by timestamps and normalized, and encoded into a high-dimensional state vector.

3. The method according to claim 1, characterized in that, The obtained interference-suppressed baseband demodulated data specifically includes: The real-time Doppler change rate and the current phase error are input to the model predictive controller to predict the optimal frequency control word and obtain the feedforward NCO drive signal. The feedforward NCO drive signal and the loop filter output are dynamically weighted and fused to obtain the NCO control word, which drives the local carrier generation to achieve predictive-corrective phase update. Based on the total interference observation value and its instantaneous frequency characteristics, the instantaneous frequency of the interference is calculated and the notch center frequency control word is generated to obtain the clean IQ baseband signal after interference anchoring. The clean IQ baseband signal is coherently demodulated and symbol decided with the local synchronization carrier to obtain interference-suppressed baseband demodulated data, and the demodulation quality index is extracted.

4. The method according to claim 1, characterized in that, The specific steps of the output waveform strategy sequence include: Based on the high-dimensional state vector, demodulation quality index, and the waveform parameters currently in use, state aggregation encoding is performed to obtain the aggregated state; The aggregated state is input into the Monte Carlo tree search decision engine, which internally maintains a virtual interference model and performs game simulation: starting from the current state, it simulates the confrontation process between our candidate waveform strategy and the virtual interference optimal interference strategy in the next K time steps. Using the bit error rate or throughput as the payoff function, it backpropagates to update the access count and cumulative payoff of each waveform action, and obtains the access probability distribution of each candidate waveform action. Random sampling is performed according to the access probability distribution, and waveform actions for the next K time steps are generated sequentially to obtain the waveform strategy sequence; The waveform strategy sequence is bound and encapsulated with the future NCO control trajectory predicted by the model predictive controller to obtain the strategy instruction package sent to the modulation layer.

5. The method according to claim 1, characterized in that, The specific steps for generating a feedforward anti-interference modulation signal include: Parse the strategy instruction package to obtain the waveform strategy sequence and the corresponding waveform parameters, and preload the waveform parameters into different memory banks of the dual-port Block RAM in chronological order to obtain the memory bank resource pool of loaded waveform parameters. The waveform index tag is inserted into the uplink source-coded data stream to be transmitted to obtain the transmitted data frame carrying the waveform index tag; The transmitted data frame is sent to the waveform switching control unit. When the waveform index tag is detected, the waveform index tag value is mapped to the high bit of the read address of the Block RAM through a pre-configured lookup table. The memory switching is completed within one clock cycle to obtain the current valid waveform parameters and output the switching completion flag. The baseband data to be modulated is digitally modulated and shaped using the current effective waveform parameters to obtain a digital intermediate frequency modulated signal; The digital intermediate frequency modulation signal is sequentially fed into the DAC module built into the RFSOC for digital-to-analog conversion, and then processed by intermediate frequency filtering, radio frequency mixing, radio frequency filtering and power amplifier to obtain an anti-interference uplink communication radio frequency signal radiated out by the antenna.

6. The method according to claim 1, characterized in that, It also includes a closed-loop feedback step: The system collects bit error rate statistics from the baseband demodulated data stream and interference power changes from the extended state observer output. The actual communication quality is then fed back to the Monte Carlo tree search decision engine to update the parameters of the virtual interference party model and the revenue function, resulting in a continuously optimized decision model.

7. The method according to claim 2, characterized in that, Of the three parallel processing channels: the high-speed path has a processing delay of less than 1 microsecond and is used for real-time demodulation and extended state observation; the medium-speed path has a processing delay of 1 to 100 microseconds and is used to generate short-time Fourier transform time-frequency plots; the low-speed path has a processing delay of more than 100 microseconds and is used to accumulate statistical features including peak-to-average power ratio and cyclic stationary spectrum.

8. The method according to claim 3, characterized in that, In the predictive-corrective Costas loop, the satellite orbit dynamics model built into the model predictive controller is used to calculate the feedforward compensation amount of the Doppler rate of change; the dynamic weighted fusion ratio of the feedforward NCO drive signal and the loop filter output is adaptively adjusted according to the current loop locking state: when the loop is unlocked, the weight of the feedforward component is greater than the weight of the feedback component, and when the loop is locked, the weight of the feedback component gradually increases.

9. The method according to claim 4, characterized in that, When the Monte Carlo tree search decision engine performs forward-looking multi-step game simulations, the simulation step number K corresponds to a time length of 20–100 milliseconds; the virtual interfering party model adopts a greedy strategy that minimizes its own gains, and the gain function is the negative logarithm of the bit error rate or a linear function of the throughput.

10. The method according to claim 5, characterized in that, The waveform index tag has a length of 8 to 16 bits; address remapping is achieved through a pre-configured read-only lookup table, the input of which is the detected waveform index tag value and the output is the high-order bits of the Block RAM read address, thus achieving interrupt-free waveform switching; the switching completion flag is used to trigger the prefetching of waveform parameters for the next memory bank.

Images