A deep learning-based super large-scale low-orbit satellite communication adaptive coding and modulation method and system

Abstract

The application discloses a kind of based on deep learning's super large scale low-orbit satellite communication adaptive coding modulation method and system, and proposes higher reliability and stability solution for super large scale low-orbit satellite earth communication system.The application first constructs low-orbit satellite earth communication simulation system and satellite-earth communication channel model;Secondly, the mobile model of single low-orbit satellite is constructed, and the distance and elevation angle change data in the process of satellite orbiting is obtained according to the simulation result;Finally, the channel state is predicted in the process of satellite movement using deep learning algorithm, and the adaptive coding modulation method is applied to low-orbit satellite communication scene.Compared with the traditional adaptive coding modulation method, the application considers the rapid change of dynamic satellite network topology, which is more suitable for real scene, and can maximize the throughput of low-orbit satellite earth communication system and improve communication efficiency based on the use of deep learning algorithm.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, and relates to the theory of ultra-large-scale low-Earth orbit satellite-to-ground communication and the design of adaptive coding and modulation schemes. Background Technology

[0002] Currently, Low Earth Orbit (LEO) satellite terrestrial networks (LSTN) have significantly enhanced the data transmission capabilities of terrestrial networks and are expected to become an important component of next-generation wireless networks (6G). Furthermore, the emergence of ultra-large-scale LEO constellations, such as SpaceX and OneWeb, has the potential to provide global communication coverage and accelerate the development of various technologies, including IoT, remote surgery, and vehicle-to-everything (V2X) communication. However, the high dynamism of ultra-large-scale LEO satellite systems introduces uncertainty in the quality of service for LSTN users, necessitating the design of more flexible communication protocols and algorithms to ensure the stability and reliability of data transmission. Currently, Starlink employs the DVB series of physical layer standards, including Constant Coding Modulation (CCM) and Adaptive Coding Modulation (ACM) methods. Traditional CCM methods are typically conservatively designed to adapt to the worst channel conditions, resulting in poor utilization of spectrum resources. Therefore, research on ACM methods based on changing channel state detection to achieve real-time switching of coding modulation schemes is crucial.

[0003] A review of existing literature reveals that much of the past research has focused on ACM schemes in data transmission. Wang et al. published a paper in 2020 titled "Adaptive modulation and coding technology in 5G system," which proposed an improved adjustment factor-optimized exponential effective signal-to-noise ratio mapping algorithm to optimize ACM technology in the downlink of the physical layer of 5G systems. Ji et al. published a paper in 2021 titled "Deep learning for adaptive modulation and coding with payload length in vehicle-to-vehicle communication systems," which proposed an ACM method using convolutional neural networks for channel state estimation. These publications all focus on data transmission in terrestrial networks with variable channel states. Xu et al. and Li et al. published articles titled "A sparse bayesian learning method of joint activity detection and channel estimation for LEO grant-free random access" at the IEEE International Workshop on Signal Processing Advances in Wireless Communications in 2023 and "A novel method of ACM optimization based on system characteristics for the Ka band TT&C and communication systems" at the IEEE Vehicular Technology Conference in 2016. These articles focus on satellite communication networks, but only on satellites and ground stations in fixed locations, without addressing satellite movement. However, in real-world scenarios, satellite positions and channel states are constantly changing, necessitating the modeling of dynamic satellite-to-ground communication scenarios and satellite channels.

[0004] Further research revealed that, in order to further optimize the ACM method in low-Earth orbit satellite communication systems, Lee et al. published a paper entitled "Neural Episodic" in IEEE Access in 2021.

[0005] Control-Based Adaptive Modulation and Coding Scheme for Inter-Satellite

[0006] The paper "Communication Link (Adaptive Modulation and Coding Scheme for Inter-Satellite Communication Link Based on Neural Condition Control)" proposes a method for adjusting the region boundary of the modulation and coding scheme based on deep reinforcement learning. This method can significantly reduce the number of paths to find the optimal MCS region boundary and use less memory. In their 2020 paper "FPGA Based Transmitter Design Using Adaptive Coding and Modulation Schemes for Low Earth Orbit Satellite Communications" published at the IEEE International Symposium on Telecommunication Technologies, Tutgun et al. proposed a high-speed transmitter based on a field-programmable gate array (FPGA) developed using ACM technology for LEO satellite communication links. These papers all consider satellite channel state estimation but neglect the potential for CQI obsolescence due to the time-varying characteristics of satellites and channels. Therefore, in addition to channel state estimation, channel state prediction is also crucial. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide an adaptive coding and modulation method and system for ultra-large-scale low-Earth orbit satellite communication based on deep learning. Through technologies such as dynamic LEO satellite communication scenario modeling, channel state estimation, channel state prediction, and MCS switching strategy, it proposes a solution with higher reliability and stability for low-Earth orbit satellite-to-ground communication systems.

[0008] Technical Solution: To achieve the above-mentioned objectives, this invention adopts the following technical solution: an adaptive coding and modulation method for ultra-large-scale low-Earth orbit satellite communication based on deep learning, comprising the following steps:

[0009] Construct a low-Earth orbit satellite-to-ground communication simulation system and a satellite-to-ground communication channel model;

[0010] A motion model of a single low-Earth orbit satellite is constructed, and the distance and elevation angle changes during the satellite's orbit are obtained based on the simulation results.

[0011] Simulations were performed on each coding and modulation scheme (MCS) under the established satellite channel environment to obtain the bit error rate (BER) - signal-to-noise ratio (SNR) curves. Based on the SNR threshold data corresponding to the preset BER threshold in the relationship curves for each MCS, a channel quality indicator (CQI) lookup table was obtained.

[0012] After the transmitting end transmits the data with the inserted pilot signal, the receiving end first estimates the SNR at this moment based on the extracted pilot signal, then uses a deep learning algorithm to predict the SNR at the next moment, maps the SNR value reflecting the channel quality to the corresponding MCS, and then feeds back the CQI to the transmitting end.

[0013] After receiving the CQI from the receiver, the sending end selects the corresponding MCS mode for the next data transmission based on the CQI lookup table.

[0014] Preferably, the Digital Video Broadcasting-Satellite 2nd Generation Extended Standard (DVB-S2X) physical layer standard is used to model the low-Earth orbit satellite-to-ground communication simulation system. The signal transmitter specifically employs BCH outer code and LDPC inner code as forward error correction coding techniques, passing through a bit interleaver before symbol mapping and modulation. The receiver performs demodulation, symbol demapping, bit deinterleaving, LDPC, and BCH decoding operations.

[0015] Preferably, the International Telecommunication Union (ITU) series of standards are used to model the satellite-to-ground communication channel. When modeling the low-Earth orbit (LEO) satellite-to-ground communication channel, only large-scale fading is considered, including free-space loss, rain attenuation prediction model (reference ITU-R P.618), cloud attenuation prediction model (reference ITU-R P.840), atmospheric attenuation prediction model (reference ITU-R P.676), and ground cover attenuation prediction model (reference ITU-R P.2108).

[0016] As a preferred method, the trajectory of a single low-Earth orbit satellite is simulated using a fixed ground receiving station to obtain data on the changes in distance and elevation angle of the satellite during its orbit around the Earth.

[0017] As a preferred method, the receiver uses a least squares (LS) estimator to estimate the channel state and obtain the SNR.

[0018] Preferably, the SNR prediction for the next time step using a deep learning algorithm is performed by using a time series prediction algorithm (AB-LSTM) that combines an autoregressive integral moving average model (ARIMA) with a bidirectional long short-term memory neural network (BiLSTM). The AB-LSTM algorithm uses BiLSTM to determine the p-order and q-order of ARIMA through the nonlinear residuals of the predicted sequence, where p represents the lag length of the time series data and q represents the lag length of the prediction error.

[0019] Preferably, for SNR prediction at time k+1, the model input {γ} k} l For an SNR time series of length l, {γ} k-l ,γ k-l+1 ,…,γ k The output of the model It is the predicted SNR sequence of length n First, iterate through the possible values ​​of p and q, then ARIMA outputs a set of n-step predictions. Then, the predicted SNR value is compared with the actual SNR value to obtain a set of residual time series {e k-n+1 ,e k-n+2 ,…,e k},in, The residual time series is set as the input to the BiLSTM neural network, and the output of the BiLSTM neural network is the predicted residual for the next time step. Find the p that minimizes the predicted residual. k and q k , will {γ k} l Set to ARIMA(p) k ,d,q k The input is , and the output is the predicted SNR value at time (k+1).

[0020] A deep learning-based adaptive coding and modulation system for ultra-large-scale low-Earth orbit satellite communication includes:

[0021] The simulation environment construction module is used to build a low-Earth orbit satellite-to-ground communication simulation system and a satellite-to-ground communication channel model; as well as to build a motion model of a single low-Earth orbit satellite, and obtain data on the changes in distance and elevation angle of the satellite during its orbit based on the simulation results;

[0022] The simulation module is used to simulate each coding and modulation scheme (MCS) in the established satellite channel environment, obtain the bit error rate (BER) - signal-to-noise ratio (SNR) curve, and obtain the channel quality indication (CQI) lookup table based on the SNR threshold data corresponding to the preset BER threshold in the relationship curve of each MCS.

[0023] The adaptive coding and modulation module is used so that after the transmitter transmits data with the inserted pilot signal, the receiver first estimates the SNR at this moment based on the extracted pilot signal, then uses a deep learning algorithm to predict the SNR at the next moment, maps the SNR value reflecting the channel quality to the corresponding MCS, and then feeds back the CQI to the transmitter; and after the transmitter receives the CQI fed back by the receiver, it selects the MCS mode corresponding to the CQI according to the CQI lookup table for data transmission at the next moment.

[0024] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of a deep learning-based adaptive coding and modulation method for ultra-large-scale low-Earth orbit satellite communication.

[0025] Beneficial effects: Compared with the prior art, the advantages of this invention are as follows: First, the deep learning-based adaptive coding and modulation method for low-Earth orbit satellite communication considers multiple channel fading factors and dynamic satellite changes when establishing low-Earth orbit satellite-to-ground communication scenarios, making it more closely aligned with real-world scenarios; Second, this invention utilizes deep learning algorithms to achieve accurate channel state prediction, and the AB-LSTM time series prediction algorithm constructed in this invention has been proven to have better prediction performance compared to baseline algorithms, capable of handling not only the linear part of the SNR sequence but also the nonlinear part; Third, the low-Earth orbit satellite communication ACM method incorporating channel state prediction effectively enhances the reliability and stability in high-latency LSTN scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a LEO satellite-to-ground communication modeling scenario according to an embodiment of the present invention.

[0027] Figure 2 This is a graph showing the loss versus frequency obtained by considering various channel attenuation factors in an embodiment of the present invention.

[0028] Figure 3 This is a diagram of the link layer mechanism in the satellite-to-ground communication process according to an embodiment of the present invention.

[0029] Figure 4 This is a flowchart of the channel state prediction algorithm used in the embodiments of the present invention.

[0030] Figure 5 This is a comparison chart of the absolute estimation errors obtained under three channel state estimation algorithms according to embodiments of the present invention.

[0031] Figure 6 This is a comparison chart of the prediction performance obtained under four channel state prediction algorithms according to the embodiments of the present invention.

[0032] Figure 7 This is the final MCS switching diagram with SNR in the embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention are described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solutions of this invention, providing detailed implementation methods and specific operating procedures. It should be understood that the specific examples described herein are merely illustrative of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0034] Example 1

[0035] The embodiments of the present invention are based on Figure 1 This paper proposes an adaptive coding and modulation method for ultra-large-scale low-Earth orbit (LEO) satellite-to-ground communication modeling scenario, based on deep learning. First, a LEO satellite-to-ground communication simulation system is modeled using the DVB-S2X physical layer standard, and a satellite-to-ground communication channel model is modeled using a series of ITU standards. Second, a motion model of a single LEO satellite is constructed, and the distance and elevation angle changes during the satellite's orbit are obtained based on the simulation results. Finally, a deep learning algorithm is used to predict the channel state during satellite motion, applying the adaptive coding and modulation method to the LEO satellite communication scenario. This method utilizes deep learning algorithms to predict channel conditions during satellite movement, applying adaptive coding and modulation methods to low-Earth orbit satellite communication scenarios. The main steps include: simulating various coding and modulation schemes (MCS) under the established satellite channel environment to obtain the bit error rate (BER) - signal-to-noise ratio (SNR) curves; and obtaining a channel quality indicator (CQI) lookup table based on the SNR threshold data corresponding to the preset BER threshold in the relationship curves for each MCS. After the transmitter transmits data with inserted pilot signals, the receiver first estimates the SNR at that moment based on the extracted pilot signals, then uses a deep learning algorithm to predict the SNR at the next moment, mapping the SNR value reflecting channel quality to the corresponding MCS, and then feeding the CQI back to the transmitter. After receiving the CQI from the receiver, the transmitter selects the corresponding MCS mode according to the CQI lookup table for data transmission at the next moment.

[0036] In the experimental scenario of this embodiment, the ground station is set as Beijing Station, and a single LEO low-Earth orbit satellite passes directly above Beijing Station. Communication with the ground station begins when the elevation angle between the satellite and the ground station is greater than 5°, at which point the communication distance between the satellite and the ground station is the greatest; the distance between the satellite and the ground station is minimized when the elevation angle is 90°. Large-scale fading is considered when modeling the satellite channel, including free space loss, atmospheric absorption loss, cloud attenuation, rain attenuation, and ground object loss. Figure 2The diagram shows the satellite channel loss. Atmospheric absorption loss is referenced to ITU-R Recommendation P.676, cloud attenuation to ITU-R Recommendation P.840, rain attenuation to ITU-R Recommendation P.618, and ground feature loss to ITU-R Recommendation P.2108. Figure 3 The diagram illustrates the link-layer mechanism in LEO satellite-to-Earth communication incorporating the ACM method. At the transmitting end, the signal undergoes BCH encoding, LDPC encoding, bit interleaving, symbol mapping, and modulation sequentially. After passing through the satellite channel, the receiving end performs demodulation, symbol demapping, bit deinterleaving, LDPC decoding, and BCH decoding. Additionally, the pilot signal is extracted and fed into the ACM module. First, channel estimation is performed based on the pilot signal. Then, the channel signal-to-noise ratio (SNR) for the next time step is predicted based on the estimated value. MCS selection is then performed based on the prediction, and the result is fed back to the transmitting end, which adjusts the encoding and modulation scheme for subsequent data. Simulations show that setting the target BER to 10... -4 The CQI lookup tables for some MCSs are shown in Table 1:

[0037] Table 1: CQI Lookup Table

[0038]

[0039]

[0040] The transmitting and receiving ends jointly maintain a CQI lookup table. After the transmitting end transmits data, the receiving end first performs SNR estimation based on the received pilot signal, then performs SNR prediction, mapping the effective SNR value reflecting the current channel quality to the CQI value, and then feeds the CQI back to the transmitting end. After receiving the CQI fed back by the terminal ground station, the transmitting end selects the transmission mode corresponding to that CQI for the next downlink data transmission according to the receiving end's CQI lookup table.

[0041] The basic objective of this embodiment is to implement high-precision channel state estimation, high-precision channel state prediction, and MCS handover strategy within the ACM module. In the channel state estimation section, this embodiment employs a least squares (LS) estimator. As a data-aided (DA) estimation method, the LS estimator has good estimation performance. Assume the pilot signal at the transmitter is x = (x1, x2, ..., x...). n If the transmitted signal x passes through channel h and is then received, the received signal can be expressed as the sum of the effects of channel h and noise n, i.e.: y = hx + n. If there are multiple pilot signals, an observation matrix X is constructed, with each row corresponding to one pilot signal. The channel is estimated by minimizing the sum of squared errors using the least squares method. The error e can be defined as: e = y - Xh. Minimizing the sum of squared errors gives the channel estimate. The expression, that is: After obtaining the channel estimate, the signal power and noise power are calculated as follows:

[0042]

[0043] The signal-to-noise ratio obtained by the estimator is given by the following formula:

[0044]

[0045] In the channel state prediction section, this embodiment adopts a time series prediction algorithm that combines the autoregressive integral moving average model (ARIMA) with the bidirectional long short-term memory neural network (BiLSTM): AB-LSTM. Figure 4 A flowchart of the AB-LSTM algorithm is provided. ARIMA can capture linear relationships in time series, while the BiLSTM algorithm avoids long-term dependency problems and captures nonlinear relationships in time series. Combining ARIMA and BiLSTM can better predict the changing channel state. Key parameters of ARIMA include the differencing order d, the autoregressive order p, and the moving average order q. For non-stationary sequences, d differencing is required to make the sequence stationary. p represents the lag length of the time series data, and q represents the lag length of the prediction error. The standard method for determining p and q is AIC, which, based on the concept of entropy, provides a standard for measuring the complexity of the prediction algorithm and the accuracy of the fitted data. The AB-LSTM algorithm uses BiLSTM to determine the p and q orders of ARIMA through the nonlinear residuals of the predicted sequence, deeply learning the nonlinear relationships contained in the SNR time series. Taking the calculation of the SNR prediction value at time k+1 as an example, the initial input of the ARIMA model {γ}... j} l For an SNR time series of length l, {γ} k-l ,γ k-l+1 ,…,γ k The output of the model It is the predicted SNR sequence of length n First, iterate through the possible values ​​of p and q, then ARIMA outputs a set of n-step predictions. Calculated using ARIMA The expression is as follows:

[0046]

[0047] Where, Δ d γ k-i For γ k-i The d-th order difference, where μ is a constant and α k-j It is the random error at time (kj), φ i (i = 1, 2, ..., p) and θ j(j=1,2,...,q) is the autocorrelation function, and p and q are the orders of ARIMA. Then, the predicted SNR values ​​are compared with the actual SNR values ​​to obtain a set of residual time series {e k-n+1 ,e k-n+2 ,…,e k},in, Set the residual time series as the input to the BiLSTM. Assume that during a forward data processing process, f k It's the Gate of Oblivion, i k It's an input gate, o k It's the output gate. The forget gate determines how many previous cell states c are represented. k-1 It will be retained. and This is the weight matrix of the forget gate, b f It is a bias term. Let be the predicted residual value at time k-1, and its expression is as follows:

[0048]

[0049] The input gate determines how much new information will be stored in the cell state. and It is the weight matrix of the input gate, b i It is a bias term; and It is the weight matrix of the candidate unit states, b c This is the bias term. First, the activation value of the input gate is calculated, and then a new candidate value is created, which can be added to the state. The expressions for the input gate and candidate cell states are as follows:

[0050]

[0051] Then, the cell state is updated based on the previous results. Represents the Hardamad product, old state C k-1 The output of the forget gate is multiplied to discard unwanted information, and the product of the input gate and the candidate value is added to update the state.

[0052]

[0053] The output of a forward process is calculated as follows: and It is the weight matrix of the output gate, b o It is a bias term.

[0054]

[0055] Considering that BiLSTM can handle reversed data, assume that the output of the reversed LSTM layer is o′k Then the output of BiLSTM is [o k ,o′ k The output of the BiLSTM neural network is the predicted residual for the next time step. Find the p that minimizes the predicted residual. k and q k Then {γ k} l Set to ARIMA(p) k ,d,q k The input is , and the output is the predicted SNR value at time (k+1). as follows:

[0056]

[0057] In the MCS selection strategy section, this embodiment adopts the fixed threshold switching method, that is, based on the predicted signal-to-noise ratio, the CQI lookup table is queried, and the MCS switching scheme is selected downwards.

[0058] Simulation experiment:

[0059] We will analyze the proposed deep learning-based adaptive coding and modulation method for ultra-large-scale low-Earth orbit satellite communication according to... Figure 1 and Figure 2 The process shown is simulated and verified. Figure 1 In the diagram, α represents the elevation angle between the satellite and the ground station, and R... e R is the Earth's radius, and h is the satellite's altitude above the Earth's surface. In our simulation, a carrier frequency of 30 GHz is used. e The distance is set to 8500km, h to 680km, and the transmission data frame length to the standard length of 64800 bits. The simulation time step is one transmission time interval (TTI). We first evaluate the performance of the designed adaptive coding and modulation algorithm from two aspects: channel state estimation algorithm and channel state prediction algorithm. Then, we perform system simulation on a satellite-to-ground communication system based on the DVB-S2X standard that adopts the adaptive coding and modulation method.

[0060] exist Figure 5 In our simulation, we obtained a comparison of the absolute errors of three channel state estimation algorithms: Minimum Mean Square Error (MMSE), Second and Fourth Moments (M2M4), and Least Squares (LS). Compared to MMSE and M2M4, LS achieves a smaller channel state estimation error. Figure 6In this paper, we compare the root mean square error (RMSE) and mean absolute error (MAE) of four channel state prediction algorithms: Autoregressive Integral Moving Average (ARIMA), Bidirectional Long Short-Term Memory (BiLSTM), Gated Unit (GRU), and AB-LSTM (a combination of ARIMA and BiLSTM). It can be seen that the AB-LSTM algorithm achieves the best channel state prediction performance compared to the other three algorithms. Figure 7 By incorporating the LS algorithm and AB-LSTM algorithm into the ACM method, we can conclude that in the satellite-to-ground communication link under the DVB-S2X standard, after performing channel state estimation and channel state prediction, the transmitter can correctly perform MCS handover at the handover threshold point as the channel SNR changes.

[0061] Example 2

[0062] This invention discloses a deep learning-based adaptive coding and modulation system for ultra-large-scale low-Earth orbit (LEO) satellite communication, comprising: a simulation environment construction module for constructing a LEO satellite-to-ground communication simulation system and a satellite-to-ground communication channel model; and a single LEO satellite motion model, obtaining distance and elevation angle change data during satellite orbiting the Earth based on the simulation results; and a simulation module for simulating various coding and modulation schemes (MCS) in the established satellite channel environment, obtaining the bit error rate (BER) - signal-to-noise ratio (SNR) curve, and pre-setting B in the relationship curve for each MCS. The SNR threshold data corresponding to the ER threshold is used to obtain a Channel Quality Indicator (CQI) lookup table. The adaptive coding and modulation module is used so that after the transmitter transmits data with inserted pilot signals, the receiver first estimates the SNR at this moment based on the extracted pilot signals, then uses a deep learning algorithm to predict the SNR at the next moment, mapping the SNR value reflecting channel quality to the corresponding MCS, and then feeding the CQI back to the transmitter. After receiving the CQI from the receiver, the transmitter selects the corresponding MCS mode for the next moment's data transmission based on the CQI lookup table. Specific implementation details are detailed in the method embodiment and will not be repeated here.

[0063] Example 3

[0064] This invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the deep learning-based adaptive coding and modulation method for ultra-large-scale low-Earth orbit satellite communication described in Embodiment 1.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based super large-scale low-orbit satellite communication adaptive coding and modulation method, characterized in that, The method comprises the following steps: A low-orbit satellite-to-ground communication simulation system and a satellite-to-ground communication channel model are constructed; A single low-orbit satellite movement model is constructed, and distance and elevation angle change data of the satellite during the process of orbiting the earth are obtained according to the simulation running result; In the established satellite channel environment, each coding and modulation scheme (MCS) is simulated to obtain a bit error rate (BER)-signal-to-noise ratio (SNR) curve, and a channel quality indicator (CQI) lookup table is obtained according to the SNR threshold data corresponding to a preset BER threshold in the relationship curve of each MCS; After the sending end transmits data with a pilot signal inserted, the receiving end first performs SNR estimation at this moment according to the extracted pilot signal, then uses a deep learning algorithm to perform SNR prediction at the next moment, maps the SNR value reflecting the channel quality to the corresponding MCS, and then feeds back the CQI to the sending end; After the sending end receives the CQI feedback from the receiving end, the MCS corresponding to the CQI is selected according to the CQI lookup table for data transmission at the next moment; The deep learning algorithm used for SNR prediction at the next moment is an autoregressive integrated moving average model (ARIMA) combined with a bidirectional long short-term memory neural network (BiLSTM) time series prediction algorithm (AB-LSTM) for SNR prediction; The AB-LSTM algorithm uses BiLSTM to determine the p-th and q-th orders of ARIMA through the nonlinear residuals of the predicted sequence, where p represents the lag length of the time series data and q represents the lag length of the prediction error; for the SNR prediction at time k+1, the model input {γ k } l For an SNR time series of length l, {γ} k-l ,γ k-l+1 ,…,γ k The output of the model It is the predicted SNR sequence of length n First, iterate through the possible values ​​of p and q, then ARIMA outputs a set of n-step predictions. Then, the predicted SNR value is compared with the actual SNR value to obtain a set of residual time series {e k-n+1 ,e k-n+2 ,…,e k },in, The residual time series is set as the input to the BiLSTM neural network, and the output of the BiLSTM neural network is the predicted residual for the next time step. Find the p that minimizes the predicted residual. k and q k , will {γ k } l Set to ARIMA(p) k ,d,q k The input is , and the output is the predicted SNR value at time (k+1).

2. The deep learning-based adaptive coding and modulation method for super large-scale low earth orbit satellite communication according to claim 1, characterized in that: The digital video broadcast-satellite second generation extension standard (DVB-S2X) physical layer standard is used to model the low-orbit satellite-to-ground communication simulation system; the sending end uses a BCH outer code and an LDPC inner code as forward error correction coding technology, passes through a bit interleaver, and then performs symbol mapping and modulation; the receiving end performs demodulation, symbol demapping, bit deinterleaving, LDPC and BCH decoding operations.

3. The deep learning-based adaptive coding and modulation method for super large-scale low earth orbit satellite communication according to claim 1, characterized in that: The International Telecommunication Union (ITU) series standards are used to model the satellite-to-ground communication channel model; when modeling the low-orbit satellite-to-ground communication channel, only large-scale fading is considered, including free space loss, a rain fade prediction model referring to ITU-R P.618, a cloud attenuation prediction model referring to ITU-R P.840, an atmospheric attenuation prediction model referring to ITU-R P.676, and a terrain loss prediction model referring to ITU-R P.2108.

4. The deep learning-based adaptive coding and modulation method for super large-scale low earth orbit satellite communication according to claim 1, characterized in that: Through a fixed ground receiving station, the movement trajectory of a single low-orbit satellite is simulated to obtain distance and elevation angle change data of the satellite during the process of orbiting the earth.

5. The deep learning-based adaptive coding and modulation method for super large-scale low earth orbit satellite communication according to claim 1, characterized in that: The receiving end uses a least square (LS) estimator to perform channel state estimation to obtain SNR.

6. A deep learning based super large scale low earth orbit satellite communication adaptive coding and modulation system, characterized in that, The method comprises: A simulation environment construction module is configured to construct a low-orbit satellite-to-ground communication simulation system and a satellite-to-ground communication channel model, and construct a single low-orbit satellite movement model to obtain distance and elevation angle change data of the satellite during the process of orbiting the earth according to the simulation running result; A simulation module is configured to simulate each coding and modulation scheme (MCS) in the established satellite channel environment to obtain a bit error rate (BER)-signal-to-noise ratio (SNR) curve, and obtain a channel quality indicator (CQI) lookup table according to the SNR threshold data corresponding to a preset BER threshold in the relationship curve of each MCS; ​ An adaptive coding and modulation module is used to transmit data with pilot signals inserted at the sending end, and the receiving end first performs SNR estimation at this moment according to the extracted pilot signals, then uses a deep learning algorithm to predict the SNR at the next moment, maps the SNR value reflecting the channel quality to the corresponding MCS, and then feeds back the CQI to the sending end; And after the sending end receives the CQI feedback from the receiving end, the MCS corresponding to the CQI is selected according to the CQI lookup table for data transmission at the next moment; the use of the deep learning algorithm to predict the SNR at the next moment is to use the time series prediction algorithm (AB-LSTM) combining the autoregressive integrated moving average model (ARIMA) and the bidirectional long short-term memory neural network (BiLSTM) to predict the SNR; The AB-LSTM algorithm uses BiLSTM to determine the p-th and q-th orders of ARIMA through the nonlinear residuals of the predicted sequence, where p represents the lag length of the time series data and q represents the lag length of the prediction error; for the SNR prediction at time k+1, the model input {γ k } l For an SNR time series of length l, {γ} k-l ,γ k-l+1 ,…,γ k The output of the model It is the predicted SNR sequence of length n First, iterate through the possible values ​​of p and q, then ARIMA outputs a set of n-step predictions. Then, the predicted SNR value is compared with the actual SNR value to obtain a set of residual time series {e k-n+1 ,e k-n+2 ,…,e k },in, The residual time series is set as the input to the BiLSTM neural network, and the output of the BiLSTM neural network is the predicted residual for the next time step. Find the p that minimizes the predicted residual. k and q k , will {γ k } l Set to ARIMA(p) k ,d,q k The input is , and the output is the predicted SNR value at time (k+1).

7. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the deep learning-based super-large-scale low-orbit satellite communication adaptive coding and modulation method according to any one of claims 1-5.

Images