Low earth orbit satellite latency jitter cancellation method and system

Abstract

The application belongs to the technical field of neural network combined with low-orbit satellite communication, and particularly relates to a low-orbit satellite delay jitter elimination method and system. The method comprises the following steps: S1, based on ephemeris assistance, acquiring a physical layer signal-to-noise ratio and a MAC layer queue depth, calculating an instantaneous elevation angle, an instantaneous Doppler frequency shift and a Doppler change rate, and extracting a historical smooth jitter feedback; S2, inputting a multi-dimensional state vector into a pre-training sequence model to predict queue backlog and end-to-end jitter under different modulation and coding schemes; S3, constructing a multi-objective optimization function, actively overdrafting a tolerable error rate upper limit in a channel rapid decline period, maintaining a high-order modulation or a high coding rate, and accelerating the release of the cache to flatten the burst queuing delay; and S4, adopting a timestamp-based leaky bucket scheduling algorithm to perform micro timing shaping on an output data stream to eliminate residual jitter caused by high-frequency oscillation of the physical layer.

Description

Technical Field

[0001] This invention belongs to the technical field of combining neural networks with low-Earth orbit satellite communication, specifically relating to a method and system for eliminating low-Earth orbit satellite delay jitter. Background Technology

[0002] Low Earth Orbit (LEO) satellite communication constellations have developed rapidly due to their low propagation latency and global coverage. LEO satellites typically orbit at altitudes between 500 km and 1500 km, with relative speeds exceeding 7 km / s compared to ground terminals. This extremely high relative speed presents two significant challenges: first, the elevation angle (EA) of the communication link changes very rapidly, leading to nonlinear and severe fluctuations in channel multipath fading and atmospheric absorption losses; second, it generates significant Doppler shift and Doppler rate of change.

[0003] Under existing adaptive modulation and coding (AMC) techniques, to cope with the rapidly deteriorating Channel State Information (CSI), the system frequently and drastically reduces the modulation order and increases coding redundancy. This precipitous drop in physical layer capacity can cause the data link layer (MAC layer) buffer queue to fill up instantly, leading to severe queuing delays and unpredictable network jitter. For deterministic, latency-sensitive services such as vehicle-to-everything (V2X) and satellite-to-ground collaborative control carried by low-Earth orbit satellites, the existing single-layer AMC mechanism has become a transmission bottleneck. Meanwhile, Long Short-Term Memory (LSTM) networks, through a coherent pipeline of forgetting unimportant information → remembering new, important information → synthesizing output, can accurately capture long-term trends and short-term fluctuations in time series.

[0004] Therefore, it is very important to design a method and system for eliminating queuing delays and network jitter caused by rapid channel changes and beam switching in low-Earth orbit satellite communications. Summary of the Invention

[0005] This invention aims to overcome the problem in existing adaptive modulation and coding technologies where a "cliff-like" drop in physical layer capacity leads to instantaneous fullness of the data link layer buffer queue, causing severe queuing delays and unpredictable network jitter. It provides a method and system for eliminating queuing delays and network jitter caused by rapid channel changes and beam switching in low-Earth orbit satellite communications.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The method for eliminating latency jitter in low-Earth orbit satellites includes the following steps:

[0008] S1, Multi-dimensional Link State Awareness:

[0009] Based on ephemeris assistance, the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer are obtained in real time. Ephemeris data of low-Earth orbit satellites are introduced to calculate instantaneous elevation angle, instantaneous Doppler frequency shift and Doppler rate of change, and historical smooth jitter feedback of the receiver is extracted.

[0010] S2, Delay Prediction:

[0011] Combining the time-space characteristics of the orbit, the multidimensional state vector obtained in step S1 is input into the pre-trained sequence model. This model uses the periodicity and determinism of the low-Earth orbit satellite orbit operation to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes during the elevation angle deterioration or beam switching window.

[0012] S3, Enhanced Modulation and Coding (E-AMC) Cross-Layer Control:

[0013] Construct a multi-objective optimization function that prioritizes minimizing network jitter and meeting hard latency constraints, while also considering the tolerable bit error rate. During the critical period when channel quality rapidly declines due to high-speed satellite movement, the E-AMC controller actively overdraws the upper limit of the service's tolerable bit error rate to maintain a high-order modulation scheme or a high coding rate, thereby accelerating the release of the MAC layer buffer and smoothing out sudden queuing delays.

[0014] S4, Transit Smoothing and Anti-Jitter Buffer Shaping:

[0015] While performing E-AMC at the physical layer, the MAC layer uses a timestamp-based leaky bucket scheduling algorithm to perform micro-time-sequence shaping on the output data stream to address the transient impacts before and after low-orbit satellite beam switching, thereby eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

[0016] Preferably, step S1, multi-dimensional link state perception, includes the following process:

[0017] First, a model of the geometric topology and kinematics based on ephemeris data is constructed. The high dynamic characteristics of low-Earth orbit satellites require the system to use ephemeris data (such as the satellite's orbital elements (TLE)) to calculate the spatial relative relationship between the satellite and the ground station in real time. Assuming the Earth is a standard sphere (or a WGS-84 ellipsoid), in the Earth-centered fixed coordinate system (ECEF), the position vector of the ground terminal is denoted as... At time t, the position vector and velocity vector of the low-orbit satellite, calculated using ephemeris, are denoted as follows: and v s Thus, the relative slant range vector between the satellite and the ground terminal can be expressed as: ;

[0018] The instantaneous physical distance between the satellite and the Earth (scalar) is: ;

[0019] Let the local horizontal plane normal vector (i.e., the zenith direction) at the location of the ground terminal be... Then the instantaneous elevation angle at time t : ;

[0020] In low-orbit communication, when When the visible elevation angle is less than the minimum specified by the system (such as 10° or 15°), the system needs to trigger beam switching or inter-satellite switching.

[0021] At the physical layer, dynamic channel and Doppler characteristic sensing is performed. The high-speed motion of low-Earth orbit satellites causes a drastic shift in carrier frequency. This shift is precisely calculated and compensated based on ephemeris and geometric topology. The instantaneous Doppler frequency shift at time t is... It is proportional to the relative radial velocity between the Earth and the space station, and is: ;

[0022] in, Where c is the center frequency of the carrier wave, and c is the speed of light. Let be the relative velocity vector of the satellite relative to the ground terminal. For predictive control in E-AMC, the Doppler rate of change... It is a key indicator for measuring the acceleration of channel degradation; its derivative yields: );

[0023] in, As a relative acceleration, near the overpass zenith, the Doppler frequency shift crosses zero, but the Doppler rate of change reaches its maximum value, which can easily cause the phase-locked loop (PLL) to lose lock and the physical layer capacity to drop sharply.

[0024] Meanwhile, the signal-to-noise ratio (SNR) SN of a low-Earth orbit (LEO) link is highly dependent on the instantaneous distance and elevation angle, and its extended link budget SNR equation is: ;

[0025] in, For free space path loss, Atmospheric absorption loss, since the signal passes through the thickness of the atmosphere and the cocutor of the elevation angle, can be approximated as follows: ( (This refers to the nominal absorption loss in the zenith direction).

[0026] At the data link layer, deep awareness of the buffer queue is implemented. At the MAC layer at the sending end, the system uses a discrete flow control model to monitor the buffer queue. Let the observation time step be... At the k-th discrete time (corresponding to physical time) ), depth of the cache queue Satisfies discrete equations: ;

[0027] in, This represents the instantaneous arrival rate (bits per second) of upper-layer business data. This represents the actual service rate (i.e., throughput) of the physical layer under the current modulation and coding scheme (MCS).

[0028] At the network layer, historical jitter is smoothed using feedback, and the receiver measures the arrival time interval of data packets. Let the transmission timestamps of the i-th data packet and the (i-1)-th data packet be respectively... and The received timestamps are respectively and Then it will shake instantly. The absolute value of the delay difference: ;

[0029] To filter out measurement noise and capture the long-term trend of network queuing, the system uses the Exponentially Weighted Moving Average (EWMA) algorithm to generate historical smooth jitter feedback. : ;

[0030] in, These are the smoothing weighting coefficients.

[0031] Then, cross-layer state tensor aggregation is performed, and the calculation results from the above four dimensions are combined to determine the system's state in each decision cycle. This aggregates highly heterogeneous parameters into a unified normalized multidimensional state tensor. As input for subsequent E-AMC prediction and decision engine: ;

[0032] The "-" above indicates that the parameter has been dimensionless or extreme value normalized to eliminate the gradient influence caused by different physical quantities on the neural network prediction. This indicates transpose.

[0033] Based on ephemeris data, the system was able to acquire the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time. It also incorporated ephemeris data from low-Earth orbit satellites to calculate the instantaneous elevation angle and Doppler rate of change, and extracted historical smoothing jitter feedback from the receiver.

[0034] As a preferred embodiment, step S2: time delay prediction, specifically includes the following process:

[0035] First, the historical state sequence window and data are preprocessed. To capture the time dependency of low-Earth orbit satellite channel degradation (such as elevation decrease and Doppler shift), the system no longer relies solely on the state at a single moment, but instead constructs a sliding time window of length L. At the decision point... Define the historical cross-level state sequence matrix for: ;

[0036] in, The normalized state tensor generated in step S1.

[0037] Then, prior fusion is performed on the LSTM-based time series feature extraction and trajectory analysis. The LSTM network is then used to process the nonlinear time series, extracting hidden features of channel fading and queue accumulation. For each time step in the sequence... In the mathematical model of the internal state update of a unit, the forget gate determines how much historical information (such as long-term orbital periodicity) is retained, as follows: ;

[0038] The control vector for the forget gate represents the shelf life or retention weight of historical states.

[0039] Its input gate and candidate memory determine how much of the new state at the current moment (such as a sudden drop in signal-to-noise ratio) is written into the core memory, as follows: ; ;

[0040] The input gate's control vector represents the adoption weight or sensitivity of the current burst state. The candidate cell state represents the core physical evolution trend extracted at the current moment.

[0041] Its cell state Updated to: ;

[0042] Its output gate and hidden state: ; ;

[0043] This is the output gate control vector, representing the exposure rate or output filter weights. It is a hidden state vector, representing the available information at the current moment. It is the Sigmoid activation function. This represents the Hadamard product (element-by-element multiplication). , , and For the trainable weight matrix of the network, , , , This is the bias vector. After L time steps of iteration, the network at the current time step... Output the final integrated hidden state vector. This vector highly condenses the current physical fading trend of the low-orbit link and the congestion trend of the MAC layer.

[0044] Then, the conditional policy mapping (E-AMC) is selected. The core of the prediction lies not only in predicting the future, but also in evaluating what the system will become if a certain modulation and coding scheme is adopted in the future. Therefore, the prediction model must be conditional on the candidate modulation and coding scheme (MCS).

[0045] Define candidate set The parameterization vector of the m-th policy in the example is: : ;

[0046] in The modulation order (e.g., 2 for QPSK, 4 for 16APSK). This is the FEC (Forward Error Correction) coding rate.

[0047] System timing characteristics With candidate strategies Perform tensor concatenation to construct policy conditional feature vectors. : ;

[0048] In the output of spatial-temporal delay and jitter prediction, the policy conditional feature vector is included. The input is fed into a decoder composed of a multilayer perceptron (MLP) to simultaneously predict the next decision period t. k+1 The core network metrics, including projected service rate prediction, are: ; (x is an arbitrary variable) is the activation function. In its expected cache queue depth prediction, considering the non-negativity of the physical queue, the output layer again introduces the ReLU activation function: ;

[0049] Then, derived latency and jitter calculations are performed. Based on the direct output of the aforementioned neural network, the expected queuing latency, a derived end-to-end indicator, is calculated using a queuing theory analytical model. ;

[0050] For extremely small positive numbers, to prevent division by zero overflow, the expected network jitter is: ;

[0051] To ensure that the neural network's predictions conform to the physical laws of low-Earth orbit satellites, the system adopts the design concept of a Physical Information Neural Network (PINN) during the offline training phase, with its total loss function... Data-driven error and physical constraint regularization term composition, ;

[0052] The data-driven error is represented by the mean squared error (MSE). ;

[0053] The physical constraint regularization term utilizes the Doppler rate of change from step S1. To improve the system's robustness to extreme channel conditions, the model is forced to output a more conservative (larger) delay prediction during periods of dramatic Doppler changes (often corresponding to the lowest elevation angle or beam switching point). The corresponding loss function is: ;

[0054] in For the sample size, Output the throughput for the i-th sample.

[0055] By combining the time-space characteristics of the orbit, the multidimensional state vector obtained in step S1 is input into the pre-trained sequence model. At the same time, by utilizing the periodicity and determinism of the low-Earth orbit satellite orbit operation, the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes during the upcoming elevation angle deterioration or beam switching window are predicted.

[0056] As a preferred approach, in step S3, the generation of the enhanced modulation and coding (E-AMC) multi-objective joint strategy, a combination of discrete multi-objective optimization, Pareto optimality, and a dynamic penalty weighting mechanism is employed to achieve an accurate solution for the optimal solution. First, a multi-dimensional optimization objective function is established, and then... From the candidate MCS (modulation coding scheme) set Select the optimal strategy Using the prediction output from step S2, a deterministic time-delay tracking target is established. (Force the system queuing delay to approach the target hard constraint set by the business) (to eliminate delay jitter) ;

[0057] This indicates the predicted queuing delay.

[0058] Next, a penalty target for exceeding the bit error rate (BER) limit was established. (A certain degree of bit error rate is allowed, but errors exceeding the business tolerance threshold are acceptable.) (Strict punishment will be imposed on certain parts): ;

[0059] At the same time, the established throughput maximization objective function (Maximize system capacity while meeting latency and bit error rate requirements): ;

[0060] Establish a physical layer state transition penalty target. (Regarding the Doppler variation rate of low-Earth orbit satellites) Its significant characteristic is that it suppresses frequent oscillations in the MCS. The strategy used in the previous cycle was m prev , (The order mapping function corresponding to strategy m): ;

[0061] Because the four objectives mentioned above inherently conflict (e.g., increasing throughput often leads to increased bit error rate penalties), there is no absolutely optimal solution that simultaneously minimizes all objectives. Therefore, the system first solves the Pareto front problem, eliminating absolutely inferior strategies. For the set... Any two strategies and A strategy is called a policy if it meets the following two conditions. Domination strategy (recorded as) For all sub-goals All have And at least one exists , making .

[0062] The controller iterates through the current candidate set. The solutions that are not dominated by any other strategy are extracted and formed into the Pareto Optimal Set. : ;

[0063] This indicates that it does not exist. express Superior ,m and This represents two different candidate strategies within the same set; that is, m represents the evaluation target that is currently defending its position. This represents the challengers who take turns challenging the winners. Through this calculation, the system can complete the task in a very short time (due to...). (Typically, the number of combinations is relatively small, usually only a few dozen) This eliminates ineffective strategies with high latency and high bit error rate, greatly reducing the search space. Then, cross-layer dynamic weight allocation and optimal MCS are performed to obtain the Pareto front. Then, the system uses a dynamic adaptive weighted sum method, combined with the MAC layer congestion status, to calculate the final comprehensive cost function and select a single optimal solution. .

[0064] System weight vector It's not static, but based on the current congestion level of the cache queue. It adjusts adaptively in real time. It considers latency weights. Introducing a non-linear activation function (such as a Sigmoid variant) when the queue depth approaches the physical upper limit Q max At that time, the latency weight increases exponentially: ;

[0065] in, Based on latency weights, The alarm threshold for the queue; and This is a constant used to control the nonlinear growth rate.

[0066] To eliminate the influence of different sub-objective dimensions, it is necessary to... The sub-objective values ​​of each strategy are normalized using Min-Max to obtain the normalized objective value. : ;

[0067] During the deterministic decision-making phase, the controller selects the optimal strategy for the next cycle from the Pareto front by solving for minimizing the comprehensive cost function. : ;

[0068] And it satisfies the normalized weight constraint: min means taking the minimum value of the summation.

[0069] When low-Earth orbit satellites experience rapid motion, resulting in a decrease in elevation angle and a sharp drop in signal-to-noise ratio, traditional mechanisms will significantly reduce their effectiveness. (throughput), resulting in A surge. However, in this invention, due to... The surge triggered With exponential amplification, the system will force the optimizer to operate with overwhelming weights. The goal is to find strategies that can accelerate queue emptying. That is, the system is willing to bear a certain error rate penalty. It also maintains high-order modulation, thus successfully "flattening" the upcoming delay peak and eliminating network jitter.

[0070] This completes the construction of a multi-objective optimization function that prioritizes minimizing network jitter and meeting hard latency constraints, while also considering the tolerable bit error rate. During the critical period when channel quality rapidly deteriorates due to high-speed satellite movement, the E-AMC controller actively overdraws the upper limit of the service's tolerable bit error rate to maintain a higher-order modulation scheme or lower coding redundancy, thereby accelerating the release of the MAC layer buffer and smoothing out sudden queuing delays.

[0071] Preferably, in step S4, transition smoothing and anti-jitter buffer shaping are performed to build a non-stationary dynamic leaky bucket controller with a pre-compensation mechanism at the MAC layer.

[0072] To eliminate micro-differences in physical layer processing time, the MAC layer shaper first assigns a strict timestamp to each data packet entering the buffer, establishing an ideal transmission timing reference.

[0073] Let the i-th data packet be at time i. The data reaches the MAC layer cache with a length of L(i). The system sets the target buffer smoothing delay based on the service type. Calculate the ideal physical layer release timestamp for this data packet. : ;

[0074] Define the microscopic arrival rate (instantaneous injection rate) of a packet sequence as: : ;

[0075] In the Adaptive Leaky Bucket of E-AMC, the traditional leaky bucket algorithm has a constant leakage rate. However, in satellite communications, the actual transmission rate at the physical layer is affected by the E-AMC strategy. Dynamic control. Therefore, the shaper needs to adjust the step-like physical layer rate provided by E-AMC. Low-pass filtering is used for smoothing to prevent "burst-pause" states in the physical layer from being directly transmitted to the business layer.

[0076] Suppose that in the discrete decision period The physical layer's deterministic service rate, based on the output of step three, is... The actual smooth leakage rate of the leak-proof tub shaper. Mathematical evolution using a first-order inertial low-pass filter: ;

[0077] in, This is a smoothing factor. The larger the value, the more stable the outflow rate, but the more MAC layer cache it will occupy; The smaller the value, the higher the degree to which it follows the high-frequency fluctuations of the physical layer.

[0078] At any time t, the instantaneous water level accumulation B(t) in the leaky bucket (MAC layer shaping buffer) follows an integral equation: ;

[0079] `max` refers to taking the maximum value, and then compensating for the transient impact of beam / satellite transit switching. Beam switching of low-Earth orbit satellites can cause physical layer interruptions on the order of hundreds of milliseconds (i.e., =0). If no intervention is taken at this point, the physical layer after the interruption is restored will "empty" the backlog queue at an extremely high burst rate, causing a huge negative jitter spike.

[0080] Assume the satellite transit handover occurs within the time window [t]. HO , t HO +T outage During this period The accumulated transient backlog during the period. for: ;

[0081] After the switch ends (t>t) HO +T outage The system enters a state with a duration of T. recovThe recovery period. The shaper forces this additional backlog to be evenly distributed over the recovery period, rather than sending it all at once. Corrected leakage rate during the recovery period. The model is as follows: ;

[0082] And it is limited by hard constraints to not exceed the maximum capacity of the current physical layer: ;

[0083] This mechanism mathematically guarantees that even after a physical interruption, the data stream can smoothly recover at a constant rate with a slight increase (rather than a step pulse), completely absorbing the transient impact brought by the transit.

[0084] Finally, the actual physical layer transmission time of the i-th data packet Determined by the leakage mechanism of the shaping device: ;

[0085] This refers to taking down the boundary.

[0086] The system calculates the actual transmission time. With ideal timestamp The deviation between them defines residual micro-jitter. : ;

[0087] Through the combined action of the E-AMC and MAC layer shapers, this invention mathematically ensures that for any data packet i, its residual micro-jitter is strictly limited within a very small tolerance range set by the system: ;

[0088] in The theoretical lower bound is in the nanosecond to microsecond range, thus completing the theoretical closed loop of "eliminating jitter and delay" across the entire link, with supremum being the supremum.

[0089] This allows the MAC layer to perform micro-time-sequence shaping on the output data stream using a timestamp-based leaky bucket scheduling algorithm to address transient impacts before and after low-orbit satellite beam switching, while simultaneously executing E-AMC at the physical layer, thus eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

[0090] The present invention also provides a low-orbit satellite delay jitter elimination system comprising:

[0091] The multidimensional link state awareness module is used to obtain the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time based on ephemeris, calculate the instantaneous elevation angle, instantaneous Doppler frequency shift and Doppler rate of change, extract the historical smooth jitter feedback of the receiver, and aggregate to generate a normalized multidimensional state tensor.

[0092] The delay prediction module is used to use a pre-trained sequence model, combined with orbital spatiotemporal features, and input the state tensor to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes.

[0093] The enhanced modulation and coding E-AMC cross-layer control module is used to construct a multi-objective optimization function. During the rapid decline of the channel, it actively overdraws the upper limit of the tolerable bit error rate to maintain the high-order modulation mode or high coding code rate and accelerates the release of the buffer.

[0094] The transit smoothing and anti-jitter buffer shaping module is used at the MAC layer to employ a timestamp-based leaky bucket scheduling algorithm to compensate for and uniformly unload transient backlogs during transit handover, thereby eliminating residual jitter.

[0095] Compared with the prior art, the beneficial effects of this invention are: (1) This invention breaks the barrier between the physical layer and the data link layer and constructs a dynamic weight system driven by "queue depth"; in the early stage of channel deterioration, if it is predicted that the MAC layer will be congested, the E-AMC controller will actively break the norm and adopt a game strategy of overdrafting part of the bit error rate margin (within the service tolerance range) to force the maintenance of high-order modulation and accelerate the emptying of the buffer; this fundamentally eliminates the macroscopic queuing jitter caused by the physical layer speed reduction and realizes the deterministic low latency guarantee of the satellite link; (2) This invention also innovatively introduces the ephemeris of low-orbit satellites (such as instantaneous elevation angle and Doppler rate of change) and adopts a conditional LSTM guided by physical laws. The neural network performs spatiotemporal prediction; the system no longer blindly follows the current signal-to-noise ratio, but can predict the channel evolution and queue accumulation trend of the next cycle in advance, realizing the proactive defense of "foresight"; this predictive scheduling makes the system extremely robust when facing the high dynamic motion of satellites; and adopts a multi-objective decision-making mechanism based on Pareto optimality, and innovatively introduces a state switching penalty term driven by the Doppler rate of change in the cost function. When the low-orbit satellite is in the period of Doppler drastic change, the system will automatically increase the cost penalty of the switching strategy, effectively suppressing the meaningless high-frequency oscillation of the modulation order, and ensuring the smoothness and stability of the underlying radio frequency hardware and physical layer processing timing; (3) The present invention innovatively designs a dynamic leaky bucket shaper with a transient integral compensation mechanism in the MAC layer, which can accurately calculate the amount of data accumulated during the switching interruption period, and in the recovery period, through the correction equation, linearly and uniformly amortize and unload these extra backlogs. This mechanism perfectly absorbs the physical layer pulse impact brought by the transit, making the violent fluctuations of the lower layer link completely "transparent" and imperceptible to the upper layer application. Attached Figure Description

[0096] Figure 1 This is a flowchart of a method for eliminating delay jitter in low-Earth orbit satellites according to the present invention; Figure 2 A Pareto optimization graph is generated for the enhanced modulation and coding (E-AMC) multi-objective strategy provided in the embodiments of the present invention. Detailed Implementation

[0097] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0098] like Figure 1 As shown, this invention discloses a method for eliminating delay jitter in low-Earth orbit satellites based on LSTM and enhanced modulation-coding, comprising the following steps:

[0099] Step 1: Multi-dimensional Link State Awareness

[0100] Ephemeris data is used to obtain the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time. Then, ephemeris data of low-Earth orbit satellites are introduced to calculate the instantaneous elevation angle and Doppler rate of change, and historical smoothing jitter feedback from the receiver is extracted.

[0101] Step 2: Delay Prediction

[0102] Combining the temporal and spatial characteristics of the orbit, the multidimensional state vector obtained in step one is input into the pre-trained sequence model. This model uses the periodicity and determinism of the low-Earth orbit satellite orbit operation to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes during the upcoming elevation angle deterioration or beam switching window.

[0103] Step 3: Enhanced Modulation and Coding (E-AMC) Cross-Layer Control

[0104] A multi-objective optimization function is constructed, prioritizing minimizing network jitter and meeting hard latency constraints while also considering the tolerable bit error rate. During the critical period when channel quality rapidly deteriorates due to high-speed satellite movement, the E-AMC controller actively overdrafts the upper limit of the service's tolerable bit error rate to maintain a higher-order modulation scheme or lower coding redundancy, thereby accelerating the release of the MAC layer buffer and smoothing out sudden queuing delays.

[0105] Step 4: Transit Smoothing and Anti-Jitter Buffer Shaping

[0106] While performing E-AMC at the physical layer, the MAC layer uses a timestamp-based leaky bucket scheduling algorithm to perform micro-time-sequence shaping on the output data stream to address the transient impacts before and after low-orbit satellite beam switching, thereby eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

[0107] The specific process of implementing the method of the present invention is as follows:

[0108] Step 1: Multi-dimensional Link State Awareness

[0109] In multi-dimensional link state awareness, the first step is to model the geometric topology and kinematics based on ephemeris data. The high dynamic characteristics of low-Earth orbit satellites require the system to use ephemeris data (such as the satellite's orbital elements (TLE)) to calculate the spatial relative relationship between the satellite and the ground station in real time. Assuming the Earth is a standard sphere (or a WGS-84 ellipsoid), in the Earth-centered fixed coordinate system (ECEF), the position vector of the ground terminal is denoted as... At time t, the position vector and velocity vector of the low-orbit satellite, calculated using ephemeris, are denoted as follows: and v sThus, the relative slant range vector between the satellite and the ground terminal can be expressed as: ;

[0110] The instantaneous physical distance between the satellite and the Earth (scalar) is: ;

[0111] Let the local horizontal plane normal vector (i.e., the zenith direction) at the location of the ground terminal be... Then the instantaneous elevation angle at time t : ;

[0112] In low-orbit communication, when When the visible elevation angle is less than the minimum specified by the system (such as 10° or 15°), the system needs to trigger beam switching or inter-satellite switching.

[0113] At the physical layer, dynamic channel and Doppler characteristic sensing is performed. The high-speed motion of low-Earth orbit satellites causes a drastic shift in carrier frequency. This shift is precisely calculated and compensated based on ephemeris and geometric topology. The instantaneous Doppler frequency shift at time t is... It is proportional to the relative radial velocity between the Earth and the space station, and is: ;

[0114] in, Where c is the center frequency of the carrier wave, and c is the speed of light. Let be the relative velocity vector of the satellite relative to the ground terminal. For predictive control in E-AMC, the Doppler rate of change... It is a key indicator for measuring the acceleration of channel degradation; its derivative yields: );

[0115] in, As a relative acceleration, near the overpass zenith, the Doppler frequency shift crosses zero, but the Doppler rate of change reaches its maximum value, which can easily cause the phase-locked loop (PLL) to lose lock and the physical layer capacity to drop sharply.

[0116] Meanwhile, the signal-to-noise ratio (SN) of a low-Earth orbit (LEO) link is highly dependent on the instantaneous distance and elevation angle, and its extended link SN budget equation is expressed as: ;

[0117] in, For free space path loss, Atmospheric absorption loss, since the signal passes through the thickness of the atmosphere and the cocutor of the elevation angle, can be approximated as follows: ( (This refers to the nominal absorption loss in the zenith direction).

[0118] At the data link layer, deep awareness of the buffer queue is implemented. At the MAC layer at the sending end, the system uses a discrete flow control model to monitor the buffer queue. Let the observation time step be... At the k-th discrete time (corresponding to physical time) ), depth of the cache queue Satisfies discrete equations: ;

[0119] in, This represents the instantaneous arrival rate (bits per second) of upper-layer business data. This represents the actual service rate (i.e., throughput) of the physical layer under the current modulation and coding scheme (MCS).

[0120] At the network layer, historical jitter is smoothed using feedback, and the receiver measures the arrival time interval of data packets. Let the transmission timestamps of the i-th data packet and the (i-1)-th data packet be respectively... and The received timestamps are respectively and Then it will shake instantly. The absolute value of the delay difference: ;

[0121] To filter out measurement noise and capture the long-term trend of network queuing, the system uses the Exponentially Weighted Moving Average (EWMA) algorithm to generate historical smooth jitter feedback. : ;

[0122] in, These are the smoothing weighting coefficients.

[0123] Then, cross-layer state tensor aggregation is performed, and the calculation results from the above four dimensions are combined to determine the system's state in each decision cycle. This aggregates highly heterogeneous parameters into a unified normalized multidimensional state tensor. As input for subsequent E-AMC prediction and decision engine: ;

[0124] The "-" above indicates that the parameter has been dimensionless or extreme value normalized to eliminate the gradient influence caused by different physical quantities on the neural network prediction. This indicates transpose.

[0125] Based on ephemeris data, the system was able to acquire the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time. It also incorporated ephemeris data from low-Earth orbit satellites to calculate the instantaneous elevation angle and Doppler rate of change, and extracted historical smoothing jitter feedback from the receiver.

[0126] Step 2: Delay Prediction

[0127] First, the historical state sequence window and data are preprocessed. To capture the time dependency of low-Earth orbit satellite channel degradation (such as elevation decrease and Doppler shift), the system no longer relies solely on the state at a single moment, but instead constructs a sliding time window of length L. At the decision point... Define the historical cross-level state sequence matrix for: ;

[0128] in, This is the normalized state tensor generated in step one.

[0129] Then, prior fusion is performed on the LSTM-based time series feature extraction and trajectory analysis. The LSTM network is then used to process the nonlinear time series to extract hidden features of channel fading and queue accumulation. For each time step in the sequence... In the mathematical model of the internal state update of a unit, its forget gate determines how much historical information (such as long-term orbital periodicity) is retained. ;

[0130] The control vector for the forget gate represents the shelf life or retention weight of historical states.

[0131] Its input gate and candidate memory determine how much of the new state at the current moment (such as a sudden drop in signal-to-noise ratio) is written into the core memory. ; ;

[0132] The input gate's control vector represents the adoption weight or sensitivity of the current burst state. The candidate cell state represents the core physical evolution trend extracted at the current moment.

[0133] Its cell state Updated to: ;

[0134] Its output gate and hidden state: ; ;

[0135] This is the output gate control vector, representing the exposure rate or output filter weights. It is a hidden state vector, representing the available information at the current moment. It is the Sigmoid activation function. This represents the Hadamard product (element-by-element multiplication). , , and For the trainable weight matrix of the network, , , , This is the bias vector. After L time steps of iteration, the network at the current time step... Output the final integrated hidden state vector. This vector highly condenses the current physical fading trend of the low-orbit link and the congestion trend of the MAC layer.

[0136] Then, the conditional policy mapping (E-AMC) is selected. The core of the prediction lies not only in predicting the future, but also in evaluating what the system will become if a certain modulation and coding scheme is adopted in the future. Therefore, the prediction model must be conditional on the candidate modulation and coding scheme (MCS).

[0137] Define candidate set The parameterization vector of the m-th policy in the example is: :

[0138] in, The modulation order (e.g., 2 for QPSK, 4 for 16APSK). This is the FEC (Forward Error Correction) coding rate.

[0139] System timing characteristics With candidate strategies Perform tensor concatenation to construct policy conditional feature vectors. : ;

[0140] In the output of spatial-temporal delay and jitter prediction, the policy conditional feature vector is included. The input is fed into a decoder composed of a multilayer perceptron (MLP) to simultaneously predict the next decision period t. k+1 Core network metrics, including expected service rate prediction, for ; (x is an arbitrary variable) is the activation function. In the prediction of the expected cache queue depth, considering the non-negativity of the physical queue, the output layer introduces the ReLU activation function again: ;

[0141] Then, derived latency and jitter calculations are performed. Based on the direct output of the aforementioned neural network, the expected queuing latency, a derived end-to-end indicator, is calculated using a queuing theory analytical model. ;

[0142] For extremely small positive numbers, to prevent division by zero overflow, the expected network jitter is: ;

[0143] To ensure that the neural network's predictions conform to the physical laws of low-Earth orbit satellites, the system adopts the design concept of a Physical Information Neural Network (PINN) during the offline training phase, with its total loss function... Data-driven error and physical constraint regularization term composition: ;

[0144] The data-driven error is represented by the mean squared error (MSE): ;

[0145] The physical constraint regularization term utilizes the Doppler rate of change from step one. To improve the system's robustness to extreme channel conditions, the model is forced to output a more conservative (larger) delay prediction during periods of dramatic Doppler changes (often corresponding to the lowest elevation angle or beam switching point). The corresponding loss function is: ;

[0146] in For the sample size, Output the throughput for the i-th sample.

[0147] By combining the time-space characteristics of the orbit, the multidimensional state vector obtained in step one is input into the pre-trained sequence model. At the same time, by utilizing the periodicity and determinism of the low-Earth orbit satellite orbit operation, the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes during the upcoming elevation angle deterioration or beam switching window are predicted.

[0148] Step 3: Enhanced Modulation and Coding (E-AMC) Cross-Layer Control

[0149] In the generation of multi-objective joint policies for Enhanced Modulation and Coding (E-AMC), a discrete multi-objective optimization method combined with Pareto optimality and a dynamic penalty weighting mechanism is employed to achieve accurate solutions for the optimal scheme. First, a multi-dimensional optimization objective function is established, and then... From the candidate MCS set Select the optimal strategy Using the predicted output from step two, a deterministic time-delay tracking target is established. (Force the system queuing delay to approach the target hard constraint set by the business) (to eliminate delay jitter) ;

[0150] The subsequent established bit error rate (BER) over-limit penalty target (A certain degree of bit error rate is allowed, but errors exceeding the business tolerance threshold BER are acceptable.) th (Strict punishment will be imposed on certain parts): ;

[0151] Then, the throughput maximization objective function was established. (Maximize system capacity while meeting latency and bit error rate requirements): ;

[0152] Establish a physical layer state transition penalty target. (Regarding the Doppler variation rate of low-Earth orbit satellites) Its significant characteristic is that it suppresses frequent oscillations in the MCS. The strategy used in the previous cycle was... , (The order mapping function corresponding to strategy m): ;

[0153] Because the four objectives mentioned above inherently conflict (e.g., increasing throughput often leads to increased bit error rate penalties), there is no absolutely optimal solution that simultaneously minimizes all objectives. Therefore, the system first solves for the Pareto Front, eliminating strategies with absolute disadvantages. For the set... Any two strategies and A strategy is called a policy if it meets the following two conditions. Domination strategy (recorded as) For all sub-goals All have And at least one exists , making .

[0154] The controller iterates through the current candidate set. The solutions that are not dominated by any other strategy are extracted and formed into the Pareto Optimal Set. : ;

[0155] This indicates that it does not exist. express Superior Here, m and m' represent two different candidate strategies within the same set; m represents the defending champion being evaluated, and m' represents the challenger taking turns. Through this calculation, the system can achieve this in a very short time (due to...). (Typically, the number of combinations is relatively small, usually only a few dozen) This eliminates ineffective strategies with high latency and high bit error rate, greatly reducing the search space. Then, cross-layer dynamic weight allocation and optimal MCS are performed to obtain the Pareto front. Then, the system uses a dynamic adaptive weighted sum method, combined with the MAC layer congestion status, to calculate the final comprehensive cost function and select a single optimal solution. .

[0156] System weight vector It's not static, but based on the current congestion level of the cache queue. It adjusts adaptively in real time. It considers latency weights. Introducing a non-linear activation function (such as a Sigmoid variant) when the queue depth approaches the physical upper limit Q max At that time, the latency weight increases exponentially: ;

[0157] in, Based on latency weights, The alarm threshold for the queue; and This is a constant used to control the nonlinear growth rate.

[0158] To eliminate the influence of different sub-objective dimensions, it is necessary to... The sub-objective values ​​of each strategy are normalized using Min-Max to obtain the normalized objective value. : ;

[0159] During the deterministic decision-making phase, the controller selects the optimal strategy for the next cycle from the Pareto front by solving for minimizing the comprehensive cost function. : ;

[0160] And it satisfies the normalized weight constraint: min means taking the minimum value of the summation.

[0161] When low-Earth orbit satellites experience rapid motion, resulting in a decrease in elevation angle and a sharp drop in signal-to-noise ratio, traditional mechanisms will significantly reduce their effectiveness. (throughput), resulting in A surge. However, in this invention, due to... The surge triggered With exponential amplification, the system will force the optimizer to operate with overwhelming weights. The goal is to find strategies that can accelerate queue emptying. That is, the system is willing to bear a certain error rate penalty. It also maintains high-order modulation, thus successfully "flattening" the upcoming delay peak and eliminating network jitter.

[0162] This completes the construction of a multi-objective optimization function that prioritizes minimizing network jitter and meeting hard latency constraints, while also considering the tolerable bit error rate. During the critical period when channel quality rapidly deteriorates due to high-speed satellite movement, the E-AMC controller actively overdraws the upper limit of the service's tolerable bit error rate to maintain a higher-order modulation scheme or lower coding redundancy, thereby accelerating the release of the MAC layer buffer and smoothing out sudden queuing delays.

[0163] Step 4: Transit smoothing and anti-jitter buffer shaping

[0164] Perform transition smoothing and anti-jitter cache shaping, and build a non-stationary Dynamic Leaky Bucket Controller (DLBC) with a pre-compensation mechanism at the MAC layer.

[0165] To eliminate micro-differences in physical layer processing time, the MAC layer shaper first assigns a strict timestamp to each packet entering the buffer, establishing an ideal transmission timing benchmark.

[0166] Let the i-th data packet be at time i. The data reaches the MAC layer cache with a length of L(i). The system sets the target buffer smoothing delay based on the service type. Calculate the ideal physical layer release timestamp for this data packet. : ;

[0167] Define the microscopic arrival rate (instantaneous injection rate) of a packet sequence as: : ;

[0168] In the adaptive leaky bucket shaper combined with E-AMC, the traditional leaky bucket algorithm has a constant leakage rate. However, in satellite communications, the actual transmission rate at the physical layer is affected by the E-AMC strategy. Dynamic control. Therefore, the shaper needs to adjust the step-like physical layer rate provided by E-AMC. Low-pass filtering is used for smoothing to prevent "burst-pause" states in the physical layer from being directly transmitted to the business layer.

[0169] Suppose that in the discrete decision period The physical layer's deterministic service rate, based on the output of step three, is... The actual smooth leakage rate of the leak-proof tub shaper. Mathematical evolution using a first-order inertial low-pass filter: ;

[0170] in, This is a smoothing factor. The larger the value, the more stable the outflow rate, but the more MAC layer cache it will occupy; The smaller the value, the higher the degree to which it follows the high-frequency fluctuations of the physical layer.

[0171] At any time t, the instantaneous water level accumulation B(t) in the leaky bucket (MAC layer shaping buffer) follows an integral equation: ;

[0172] `max` refers to taking the maximum value, and then compensating for the transient impact of beam / satellite transit switching. Beam switching of low-Earth orbit satellites can cause physical layer interruptions on the order of hundreds of milliseconds (i.e., =0). If no intervention is taken at this point, the physical layer after the interruption is restored will "empty" the backlog queue at an extremely high burst rate, causing a huge negative jitter spike.

[0173] Assume the satellite transit handover occurs within the time window [t]. HO , t HO +T outage During this period The accumulated transient backlog during the period. for: ;

[0174] After the switch ends (t>t) HO +T outage The system enters a state with a duration of T. recov The recovery period. The shaper forces this additional backlog to be evenly distributed over the recovery period, rather than sending it all at once. Corrected leakage rate during the recovery period. The model is as follows: ;

[0175] And it is limited by hard constraints to not exceed the maximum capacity of the current physical layer: ;

[0176] This mechanism mathematically guarantees that even after a physical interruption, the data stream can smoothly recover at a constant rate with a slight increase (rather than a step pulse), completely absorbing the transient impact brought by the transit.

[0177] Finally, the actual physical layer transmission time of the i-th data packet Determined by the leakage mechanism of the shaping device: ;

[0178] inf refers to the infimum.

[0179] The system calculates the actual transmission time. With ideal timestamp The deviation between them defines residual micro-jitter. : ;

[0180] Through the combined action of the E-AMC and MAC layer shapers, this invention mathematically ensures that for any data packet i, its residual micro-jitter is strictly limited within a very small tolerance range set by the system: ;

[0181] in The theoretical lower bound is in the nanosecond to microsecond range, thus completing the theoretical closed loop of "eliminating jitter and delay" across the entire link, with supremum being the supremum.

[0182] This allows the MAC layer to perform micro-time-sequence shaping on the output data stream using a timestamp-based leaky bucket scheduling algorithm to address transient impacts before and after low-orbit satellite beam switching, while simultaneously executing E-AMC at the physical layer, thus eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

[0183] In addition, the present invention also provides a low-orbit satellite delay jitter elimination system comprising:

[0184] The multidimensional link state awareness module is used to obtain the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time based on ephemeris, calculate the instantaneous elevation angle, instantaneous Doppler frequency shift and Doppler rate of change, extract the historical smooth jitter feedback of the receiver, and aggregate to generate a normalized multidimensional state tensor.

[0185] The delay prediction module is used to use a pre-trained sequence model, combined with orbital spatiotemporal features, and input the state tensor to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes.

[0186] The enhanced modulation and coding E-AMC cross-layer control module is used to construct a multi-objective optimization function. During the rapid decline of the channel, it actively overdraws the upper limit of the tolerable bit error rate to maintain the high-order modulation mode or high coding code rate and accelerates the release of the buffer.

[0187] The transit smoothing and anti-jitter buffer shaping module is used at the MAC layer to employ a timestamp-based leaky bucket scheduling algorithm to compensate for and uniformly unload transient backlogs during transit handover, thereby eliminating residual jitter.

[0188] like Figure 2 As shown, this invention utilizes the above method to perform Pareto optimization on the enhanced modulation and coding (E-AMC) multi-objective strategy. The Y-axis (Objective 1: Delay Degradation): the deviation between the expected queuing delay and the target delay. A larger deviation indicates a higher probability of congestion and network jitter. The X-axis (Objective 2: Bit Error Degradation): the penalty for the bit error rate exceeding the system's tolerance threshold. The further to the right, the greater the sacrifice in reliability.

[0189] exist Figure 2 In this scenario, we assume the low-Earth orbit satellite is in a window of moderate rain attenuation and rapid changes in Doppler shift. Simultaneously, the MAC layer at the transmitter is experiencing rapid queue backlog due to a decrease in physical layer capacity in the previous second. The target queuing delay is T. target =20ms. To guarantee absolutely no packet loss (bit error rate of 0), low-order modulation (such as QPSK) must be used, but this will cause data to fail to be transmitted, with the delay soaring to over 50ms. If high-order modulation (such as 16APSK) is forcibly used to transmit data per second, the delay is reduced, but the bit error rate will exceed the limit. The controller predicts 9 alternative modulation and coding schemes (MCS1 to MCS9), calculates the "delay penalty" and "bit error penalty" they may produce in the next second, and marks them on the coordinate system. The system performs Pareto-dominated filtering: observation Figure 2(MCS7, MCS8, MCS9). Taking MCS7 as an example, its bit error rate penalty is higher than MCS3, and its latency penalty is higher than MCS4. Mathematically, MCS7 is completely dominated by other strategies. The remaining points (MCS1, MCS2, MCS3, MCS4, MCS5, MCS6) form the red Pareto Front. On this solid line, no strategy can simultaneously reduce latency and bit error rate; reducing one must come at the expense of the other. The rigid choice of traditional AMC: Traditional systems rigidly adhere to the solid line of "bit error rate must be acceptable," thus choosing the top-left MCS1 or MCS2 (X-axis penalty is 0). The cost is a latency penalty of 20-35ms on the Y-axis, directly causing severe network jitter and stuttering at the service layer. The intelligent game theory of this invention's E-AMC: Because the system senses a backlog in the MAC layer queue, it triggers an exponential surge in the latency weight w1. The optimizer glides along the red line to the lower right to find the point with the lowest total cost, and finally locks in MCS4 (8PSK 3 / 4).

[0190] In summary, this invention discloses a method and system for eliminating latency jitter in low-Earth orbit satellites based on LSTM and enhanced modulation-coding. This invention breaks down the barriers of traditional network protocol stacks, proposing a bottom-up holographic perception and top-down active control cross-layer AI scheduling closed-loop system. In the ephemeris-assisted multi-dimensional state perception (cross-layer aggregation) process, it no longer relies solely on signal-to-noise ratio (SNR), but integrates stellar geometry and network features. The system introduces satellite ephemeris calculations to determine instantaneous elevation angle and Doppler rate of change, and simultaneously extracts the physical layer SNR, MAC layer cache queue depth, and network layer historical jitter, aggregating them into a multi-dimensional normalized state tensor. Then, a conditional LSTM is used for spatiotemporal latency prediction (active defense), inputting the state tensor into a Conditional Long Short-Term Memory (CLSTM) network that incorporates physical laws. The model not only extracts orbital fading periodic features but also uses candidate alternative strategies (MCS) as conditional inputs, accurately predicting how network congestion and jitter will evolve in the next second if a certain strategy is adopted. Then, Pareto multi-objective E-AMC is used for joint decision-making (core game), constructing a multi-objective cost function that includes "hard latency constraints," "tolerable bit error rate," "maximum throughput," and "state switching penalty." When severe congestion is predicted, the dynamic weighting mechanism exponentially amplifies latency priority, prompting the system to proactively choose a strategy of "overdrafting part of the bit error rate margin and forcibly maintaining high-order modulation" in the Pareto front, thereby accelerating queue emptying and directly smoothing out latency peaks at the physical layer. Finally, a transit smoothing shaping based on dynamic leaky buckets (micro-jitter reduction) is performed. For physical layer hard interruptions caused by low-Earth orbit satellite beam switching, a dynamic shaper with a "transient integral compensation equation" is designed at the MAC layer. This accurately captures the backlogged data during the interruption and linearly and uniformly unloads it during the recovery period, completely absorbing the underlying pulse impacts brought by the message storm. This allows the MAC layer to perform micro-time-sequence shaping on the output data stream using a timestamp-based leaky bucket scheduling algorithm to address transient impacts before and after low-orbit satellite beam switching, while simultaneously executing E-AMC at the physical layer, thus eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

[0191] The innovation of this invention lies in: introducing ephemeris geometric features, using the strong periodicity and determinism of satellite operation to predict channel degradation in advance by modeling instantaneous elevation angle and Doppler rate of change; a physically guided loss function, introducing the idea of ​​physical information neural networks into model training, using Doppler drastic changes as a penalty term, forcing the model to output conservative delay estimates under extreme conditions; and in the conditional policy mapping, the model not only predicts the environment, but also evaluates from a forward-looking perspective how a certain modulation and coding scheme (MCS) would be adopted, and predicts how future congestion would evolve, providing a time window for proactive jitter defense.

[0192] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for eliminating delay jitter in low-orbit satellites, characterized in that, Includes the following steps: S1, Multi-dimensional Link State Awareness: Based on ephemeris assistance, the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer are obtained in real time. Ephemeris data of low-Earth orbit satellites are introduced to calculate instantaneous elevation angle, instantaneous Doppler frequency shift and Doppler rate of change, and historical smooth jitter feedback of the receiver is extracted. S2, Delay Prediction: Combining the time-space characteristics of the orbit, the multidimensional state vector obtained in step S1 is input into the pre-trained sequence model. This model uses the periodicity and determinism of the low-Earth orbit satellite orbit operation to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes during the elevation angle deterioration or beam switching window. S3, Enhanced Modulation and Coding (E-AMC) Cross-Layer Control: Construct a multi-objective optimization function that prioritizes minimizing network jitter and meeting hard latency constraints, while also considering the tolerable bit error rate. During the critical period when channel quality rapidly declines due to high-speed satellite movement, the E-AMC controller actively overdraws the upper limit of the service's tolerable bit error rate to maintain a high-order modulation scheme or a high coding rate, thereby accelerating the release of the MAC layer buffer and smoothing out sudden queuing delays. S4, Transit Smoothing and Anti-Jitter Buffer Shaping: While performing E-AMC at the physical layer, the MAC layer uses a timestamp-based leaky bucket scheduling algorithm to perform micro-time-sequence shaping on the output data stream to address the transient impacts before and after low-orbit satellite beam switching, thereby eliminating residual jitter caused by high-frequency oscillations in the physical layer state.

2. The method for eliminating low-orbit satellite delay jitter according to claim 1, characterized in that, In step S1, the multi-dimensional link state awareness includes the process of constructing a cross-layer system state tensor, as follows: Calculate the instantaneous elevation angle between the satellite and the ground terminal using ephemeris data. Instantaneous Doppler frequency shift and Doppler rate of change : ； ； )； in, This represents the relative slant range vector between the satellite and the ground terminal. Represents the local horizontal plane normal vector at the location of the ground terminal; Where c is the center frequency of the carrier wave, and c is the speed of light; This represents the relative velocity vector of the satellite relative to the ground terminal. The instantaneous elevation angle, instantaneous Doppler frequency shift, and Doppler rate of change are compared with the physical layer signal-to-noise ratio. Buffer queue depth in the evolution of discrete flow control models And historical jitter feedback after exponential weighted smoothing. The extreme value normalization process is performed, and the data is aggregated into a state tensor. : ； in, This represents the k-th discrete decision time; "-" indicates that the corresponding parameters have been dimensionless or extreme value normalized.

3. The method for eliminating low-orbit satellite delay jitter according to claim 2, characterized in that, In step S2, the sequence model is a Conditional Long Short-Term Memory (LSTM) network with introduced physical constraints, as follows: This model uses the temporal hidden states extracted by LSTM. With candidate modulation and coding strategy parameterization vector Perform tensor concatenation to construct policy conditional feature vectors. The expected service rate is obtained by decoding through a multilayer perceptron. and expected cache queue depth ; The loss function during model training includes data-driven error. and physical constraint regularization term ; Physical constraint regularization term The Doppler rate of change drives the model to output a conservative time-delay prediction, where For the sample size, Output the throughput for the i-th sample.

4. The method for eliminating low-orbit satellite delay jitter according to claim 3, characterized in that, In step S3, the multi-objective optimization function includes: Deterministic delay tracking target ;in, To anticipate queuing delays, Hard constraints on business objectives; Error rate exceeding the limit penalty target ;in, Let m be the expected bit error rate when using strategy m. The error rate threshold tolerable by the business; Throughput maximization target ; Physical layer state switching penalty target ; It is a modulation order mapping function. The strategy used in the previous cycle.

5. The method for eliminating low-orbit satellite delay jitter according to claim 4, characterized in that, In step S3, Pareto optimality is used to screen candidate strategies, and the specific process is as follows: Traverse the set of candidate strategies Eliminating dominated strategies yields the Pareto front: ； in, This indicates that it does not exist. express Superior , m and This represents two different candidate strategies within the same set; that is, m represents the evaluation target that is currently defending its position. This refers to challengers who take turns competing. It is a Pareto non-dominated solution set; Based on the current cache queue depth The time delay weight is dynamically calculated using a nonlinear activation function. : ； in, Based on latency weights, The alarm threshold for the queue; and All of these are constants that control the nonlinear growth rate; This represents the current congestion level of the cache queue; Solving for the optimal policy after applying Min-Max normalization to the policy in the Pareto front. ; ； in, This is the normalization target value.

6. The method for eliminating low-orbit satellite delay jitter according to claim 5, characterized in that, In step S4, the leaky bucket scheduling algorithm includes dynamic leaky bucket shaping and transit handover compensation, the specific process of which is as follows: Set target buffer smooth delay Calculate the ideal physical layer release timestamp for each data packet. ; ； in, This represents the time when the i-th data packet arrives at the MAC layer buffer. The smoothed leakage rate is calculated using a first-order inertial low-pass filter. : ； in, For deterministic service rates; , is the smoothing factor; Accumulated transient backlog during transit handover interruption : ； in, The start time of the border crossing handover. The duration of the physical layer interruption. This refers to the instantaneous arrival rate of data packets. During the recovery period Evenly amortize to leakage rate : ； in This indicates the actual smooth leakage rate of the leak-proof tub shaper. The duration of the recovery period; Actual data packet sending time ; in, This refers to the length of the data packet. For ideal timestamps; inf refers to the infimum; It is a time variable; Setting residual micro-jitter : ； ； in, The lower bound of the theory is in the nanosecond to microsecond range, while sup refers to the supremum.

7. A low-Earth orbit satellite delay jitter elimination system, used to implement the low-Earth orbit satellite delay jitter elimination method according to any one of claims 1-6, characterized in that, The low-orbit satellite delay jitter elimination system includes: The multidimensional link state awareness module is used to obtain the instantaneous signal-to-noise ratio of the physical layer and the buffer queue depth of the MAC layer in real time based on ephemeris, calculate the instantaneous elevation angle, instantaneous Doppler frequency shift and Doppler rate of change, extract the historical smooth jitter feedback of the receiver, and aggregate to generate a normalized multidimensional state tensor. The delay prediction module is used to use a pre-trained sequence model, combined with orbital spatiotemporal features, and input the state tensor to predict the queue backlog and end-to-end jitter that may occur under different alternative modulation and coding schemes. The enhanced modulation and coding E-AMC cross-layer control module is used to construct a multi-objective optimization function. During the rapid decline of the channel, it actively overdraws the upper limit of the tolerable bit error rate to maintain the high-order modulation mode or high coding code rate and accelerates the release of the buffer. The transit smoothing and anti-jitter buffer shaping module is used at the MAC layer to employ a timestamp-based leaky bucket scheduling algorithm to compensate for and uniformly unload transient backlogs during transit handover, thereby eliminating residual jitter.

Images