A method and system for suppressing phase noise in satellite communication systems based on adversarial networks

Abstract

This invention provides a method and system for suppressing phase noise in a satellite communication system based on adversarial networks, comprising: performing frequency compensation on a noisy satellite communication signal to obtain a frequency-compensated signal; acquiring a noisy signal feature vector corresponding to the noisy satellite communication signal; inputting the noisy signal feature vector into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term time-series dependency data of the phase noise to obtain a first phase compensation amount; and using the first phase compensation amount to suppress the phase noise in the frequency-compensated signal to obtain a denoised satellite communication signal, thereby achieving suppression of high-order QAM phase noise. This allows for learning complex time-varying noise distributions, improving dynamic tracking capabilities, and more accurately suppressing high-order QAM phase noise.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method and system for suppressing phase noise in a satellite communication system based on adversarial networks. Background Technology

[0002] Noise in communication systems is mainly divided into additive noise and multiplicative noise. Additive noise is represented by additive white Gaussian noise in the channel, while multiplicative noise is represented by phase noise. From the perspective of the frequency source, phase noise represents the frequency stability of these devices. Long-term stability is mainly caused by slow frequency changes due to factors such as temperature and aging; short-term stability refers to random and rapid phase or frequency fluctuations.

[0003] Ideally, a single-frequency signal source should have a spectrum consisting of an infinitely narrow spectral line. In practical applications, however, the spectrum of any signal is never absolutely pure; the measurable spectral line will have a certain width and will contain periodic spurious interference or random phase shifts, which can be termed phase noise. In practical systems, phase noise primarily originates from frequency processing modules such as local oscillators, modulators, demodulators, frequency dividers, mixers, up-converters, and so on. Other factors, such as the nonlinearity of RF modules, multipath fading and Doppler effects in the channel, and sampling frequency deviations in the system, also introduce phase noise. From a frequency domain perspective, the main component of phase noise is the noise located far from the carrier frequency, primarily the phase noise of the voltage-controlled oscillator (VCO); the range closer to the carrier frequency is the phase noise of the reference signal source.

[0004] In satellite communication systems, due to changes in the space environment and the large dynamics of satellite motion speed, phase noise is dynamic, exhibiting time-varying, nonlinear, and memory characteristics, making it difficult for traditional linear models to accurately model.

[0005] Phase noise in the local oscillator signal can affect carrier frequency tracking accuracy and increase the system bit error rate (BER). Higher modulation orders have lower tolerance for phase noise. High-order QAM (such as 256QAM and 1024QAM) is extremely sensitive to phase noise; even a small phase shift can lead to significant increases in inter-symbol interference (ISI) and BER. Existing compensation algorithms based on pilot or decision feedback suffer from insufficient convergence speed in high-dynamic scenarios and incur significant pilot overhead, impacting spectral efficiency.

[0006] Therefore, how to improve the suppression capability of high-order QAM phase noise is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention provides a phase noise suppression method and system for satellite communication systems based on adversarial networks, which addresses the shortcomings of existing technologies such as poor suppression capability of high-order QAM phase noise, easy failure to converge, large pilot overhead, and impact on spectral efficiency.

[0008] On one hand, the present invention provides a phase noise suppression method for satellite communication systems based on adversarial networks, comprising:

[0009] Frequency compensation is performed on the noisy satellite communication signal to obtain a frequency-compensated signal;

[0010] Obtain the noisy signal feature vector corresponding to the noisy satellite communication signal;

[0011] The noisy signal feature vector is input into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term temporal dependence data of phase noise, thereby obtaining the first phase compensation amount.

[0012] Using the first phase compensation amount, the phase noise in the frequency-compensated signal is suppressed to obtain a noise-reduced satellite communication signal.

[0013] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the training process of the generator includes:

[0014] Construct a generator network and a generator loss function, wherein the generator loss function includes multiple first loss terms;

[0015] The first loss gradient of each first loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each first loss term.

[0016] Based on each first loss gradient, the total loss gradient of the generator loss function is obtained;

[0017] The generator network's parameters are updated by backpropagating using the total loss gradient of the generator's loss function until the convergence condition is met, thus obtaining the generator.

[0018] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the generator loss function includes a mean square error loss term;

[0019] Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including:

[0020] The maximum likelihood algorithm is used to calculate the actual required compensation amount based on the received data, transmitted data, frame synchronization length, and pilot block length in the known sequence after frame synchronization is completed.

[0021] Based on the actual demand compensation amount and the second phase compensation amount, the mean square error of the mean square error loss term is calculated;

[0022] The mean square error loss gradient of the mean square error loss term is obtained by differentiating the second phase compensation amount using the mean square error.

[0023] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the generator loss function includes a single-sideband power spectrum loss term;

[0024] Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including:

[0025] The phase power spectrum of the second phase compensation is calculated to obtain the estimated phase noise power spectrum;

[0026] Based on the estimated phase noise power spectrum and the prior phase noise power spectrum, the single-sideband power spectrum loss value of the single-sideband power spectrum loss term is calculated.

[0027] The single-sideband power spectrum loss gradient of the single-sideband power spectrum loss term is obtained by differentiating the second phase compensation amount with the single-sideband power spectrum loss value.

[0028] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the generator loss function includes a bit error rate enhancement loss term;

[0029] Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including:

[0030] Acquire padding frame data, demodulation soft information during demodulation, and re-encoding verification data after decoding from the known sequence after frame synchronization is completed;

[0031] The bit error rate enhancement loss value is calculated using the padding frame data, the demodulation soft information, and the recoding check data.

[0032] The bit error rate enhancement loss gradient of the bit error rate enhancement loss term is obtained by differentiating the second phase compensation amount with the bit error rate enhancement loss value.

[0033] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the generator loss function includes a total variation regularization loss term;

[0034] Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including:

[0035] The total variation canonical loss value is calculated using the second phase compensation amount and the third phase compensation amount at adjacent time points.

[0036] The total variation regularization loss gradient of the total variation regularization loss term is obtained by differentiating the second phase compensation amount with the total variation regularization loss value.

[0037] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the generator loss function includes an adversarial loss term;

[0038] Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including:

[0039] Using a discriminator, the confidence probability of the noise-reduced satellite communication signal during the current training is calculated based on the ideal noise-free satellite communication signal and the noise-reduced satellite communication signal during the current training.

[0040] Based on the confidence probability, calculate the loss value of the adversarial loss term;

[0041] The loss gradient of the adversarial loss term is obtained by differentiating the second phase compensation amount with the loss value of the adversarial loss term.

[0042] According to the present invention, a phase noise suppression method for a satellite communication system based on adversarial networks is provided, wherein the training process of the discriminator includes:

[0043] Construct a discriminant network and a discriminator loss function, wherein the discriminator loss function includes multiple second loss terms;

[0044] The second loss gradient of each second loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each second loss term.

[0045] Based on each second loss gradient, the total loss gradient of the discriminator loss function is obtained;

[0046] Backpropagation is performed using the total loss gradient of the discriminator loss function to update the parameters of the discriminator network until the convergence condition is met, thus obtaining the discriminator.

[0047] The phase noise suppression method for a satellite communication system based on adversarial networks provided by the present invention further includes:

[0048] The ratio of the expected contribution magnitude of the target loss term among the plurality of first loss terms to the initial weight coefficient of the target loss term is set as the actual weight of the target loss term; or

[0049] Based on the task objective, the importance of the target loss item is determined, and based on the preset correlation between importance and adjustment coefficient, a first adjustment coefficient corresponding to the importance of the target loss item is determined. The current weight of the target loss item is then adjusted to obtain its actual weight; or

[0050] Obtain the validation results of the validation set during each training process, and based on the preset correlation between the validation results and the adjustment coefficient, determine the second adjustment coefficient corresponding to the validation results, adjust the current weight of the target loss term, and obtain the actual weight of the target loss term; or

[0051] Obtain the signal characteristics of the sample signal, and set the actual weight of the target loss term based on the signal characteristics; or

[0052] The actual weight of the target loss term is dynamically adjusted by monitoring the gradient magnitude of the target loss term.

[0053] On the other hand, the present invention also provides a dynamic phase noise suppression device for a high-order satellite communication system based on adversarial networks, comprising:

[0054] The first compensation module is used to perform frequency compensation on noisy satellite communication signals to obtain a frequency-compensated signal.

[0055] The acquisition module is used to acquire the noisy signal feature vector corresponding to the noisy satellite communication signal;

[0056] The second compensation module is used to input the noisy signal feature vector into a generator pre-trained based on an adversarial network, extract local fluctuation data and long-term time-series dependency data of phase noise, and obtain the first phase compensation amount.

[0057] The suppression module is used to suppress the phase noise in the frequency-compensated signal using the first phase compensation amount, so as to obtain a noise-reduced satellite communication signal.

[0058] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the phase noise suppression method for a satellite communication system based on adversarial networks as described above.

[0059] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the phase noise suppression method for satellite communication systems based on adversarial networks as described above.

[0060] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the phase noise suppression method for satellite communication systems based on adversarial networks as described above.

[0061] The present invention provides a satellite communication system phase noise suppression method and system based on adversarial networks. This method involves: frequency compensation of noisy satellite communication signals to obtain a frequency-compensated signal; acquisition of the noisy signal feature vector corresponding to the noisy satellite communication signal; inputting the noisy signal feature vector into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term time-series dependency data of the phase noise, obtaining a first phase compensation amount; and using the first phase compensation amount to suppress the phase noise in the frequency-compensated signal to obtain a denoised satellite communication signal. This achieves suppression of high-order QAM phase noise, enabling the learning of complex time-varying noise distributions, improving dynamic tracking capabilities, and more accurately suppressing high-order QAM phase noise. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0063] Figure 1 This is a flowchart illustrating the phase noise suppression method for a satellite communication system based on adversarial networks provided in an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the generator training process;

[0065] Figure 3 This shows the receiver constellation diagram effect under different damage conditions;

[0066] Figure 4 When there is no noise, the phase noise suppression effect of the two schemes;

[0067] Figure 5 The phase noise suppression effects of the two schemes are compared when there is noise.

[0068] Figure 6This is a schematic diagram of the structure of the dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks provided in an embodiment of the present invention;

[0069] Figure 7 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0071] Figure 1 This is a flowchart illustrating the phase noise suppression method for satellite communication systems based on adversarial networks provided in this embodiment of the invention.

[0072] like Figure 1 As shown, the execution subject of the phase noise suppression method for satellite communication systems based on adversarial networks provided in this embodiment of the invention can be an electronic device, and the method mainly includes the following steps:

[0073] 101. Perform frequency compensation on noisy satellite communication signals to obtain frequency-compensated signals;

[0074] In a specific implementation process, noisy satellite communication signals can be synchronized and frequency offset corrected to complete frequency compensation, eliminate carrier frequency offset, and obtain a frequency-compensated signal.

[0075] 102. Obtain the noisy signal feature vector corresponding to the noisy satellite communication signal;

[0076] In a specific implementation process, feature extraction can be performed on the frequency-compensated signal to obtain phase and amplitude estimates. During frequency compensation, Doppler shift estimates can be obtained based on orbit predictions and blind data estimations. Signal-to-noise ratio (SNR) estimates can be obtained based on pilot signals and frame synchronization codes from known sequences. These phase, amplitude, Doppler shift, and SNR estimates are then combined to form a noisy signal feature vector corresponding to the noisy satellite communication signal. Here, it is assumed that the frequency-compensated signal is... Then the phase estimate is: The magnitude is estimated as follows: .in, For the real part, This is the imaginary part.

[0077] 103. Input the noisy signal feature vector into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term temporal dependence data of phase noise to obtain the first phase compensation amount;

[0078] In a specific implementation, a generator is pre-trained, and the generator is used to simulate a dynamic phase noise model. The feature vectors of the noisy signal are jointly extracted to obtain the first phase compensation amount. The specific function of the dynamic phase noise model is as follows (1):

[0079] (1)

[0080] in, This represents the oscillator phase noise, which typically follows a Wiener process: , This represents Gaussian white noise.

[0081] This represents the phase shift caused by the Doppler frequency shift, which is the Doppler frequency shift. The integral value, i.e. .

[0082] This indicates the phase disturbance caused by the time-varying characteristics of the channel.

[0083] In a specific implementation, dynamic phase noise modeling and suppression can be achieved using a generative adversarial network (GAN). This GAN can include a generator and a discriminator. The generator simulates the time-varying characteristics of dynamic phase noise to generate a phase-compensated signal. The discriminator distinguishes the compensated signal from the ideal noise-free satellite communication signal, guiding the generator to optimize the compensation effect.

[0084] In a specific implementation, the generator training process may include:

[0085] (1) Construct the generator network and generator loss function;

[0086] In a specific implementation, to achieve the dynamic simulation characteristics of the phase, the generator network can employ a CNN+LSTM hybrid network to jointly extract local fluctuations in the phase noise. ) and long-term time dependence ( and The CNN layer is used to extract local phase fluctuation features. It consists of three convolutional layers with a kernel size of 3×3. The LSTM layer (64 hidden units) is used to capture the temporal correlation of phase noise. The CNN output is flattened into a time series and used as the input to the LSTM. During training, the input of this generative network is the feature vector of the sample noisy signal, and the output is the phase compensation amount corresponding to the feature vector of the sample noisy signal. The phase compensation amount output after each training iteration is recorded as the second phase compensation amount for that training iteration.

[0087] In a specific implementation, the generator loss function includes multiple first loss terms, which include a mean squared error loss term, a single-sideband power spectrum loss term, a bit error rate enhancement loss term, a total variation regularization loss term, and an adversarial loss term.

[0088] Specifically, the generator loss function can be found in equation (2):

[0089] (2)

[0090] In the above formula, To counteract the loss term, G represents the generator network, D represents the discriminator network, and x is a randomly input sample of noisy satellite communication signal. The mean squared error loss term is denoted as... , The single-sideband power spectral loss term is denoted as... , For the bit error rate enhancement loss term, The total variation regularized loss term is denoted as... . The weights for the corresponding loss terms are given by k, where k represents the k-th time.

[0091] (2) Differentiate each first loss term with respect to the second phase compensation amount obtained in the current training to obtain the first loss gradient of each first loss term;

[0092] In a specific implementation, the mean squared error loss term can be calculated by the following method:

[0093] a1. Using the maximum likelihood algorithm, and based on the received data, transmitted data, frame synchronization length, and pilot block length in the known sequence after frame synchronization is completed, calculate the actual required compensation amount.

[0094] Specifically, after each training and phase noise compensation, signal demodulation can be performed to complete frame synchronization, decoding, and other operations. After frame synchronization, the frame synchronization header can be considered known data. Therefore, the actual required compensation amount can be calculated by combining the received data, transmitted data, frame synchronization length, and pilot block length, thus improving the estimation accuracy of the actual required compensation amount. The received data includes received pilots and / or frame header symbols, and the transmitted data includes transmitted pilots and / or frame header symbols. Specifically, the actual required compensation amount can be calculated according to equation (3):

[0095] (3)

[0096] in, This refers to the i-th received symbol in the received data, which can be either a received pilot symbol or a frame header symbol. This refers to the i-th transmitted symbol in the transmitted data, which can be either a pilot symbol or a frame header symbol. It is the sum of the frame synchronization length and the pilot block length. For example, if there are 32 pilot symbols and 16 frame header symbols, then... The length is 48.

[0097] b1. Based on the actual demand compensation amount and the second phase compensation amount, calculate the mean square error of the mean square error loss term;

[0098] For the specific calculation formula, please refer to the content of the mean square error loss term in equation (2), which will not be repeated here.

[0099] c1. Differentiate the second phase compensation amount using the mean square error to obtain the mean square error loss gradient of the mean square error loss term.

[0100] Mean square error loss term for second phase compensation The gradient is given by equation (4):

[0101] (4)

[0102] In a specific implementation, the single-sideband power spectrum This is a commonly used representation of phase noise in engineering or measurement. Single-sideband power spectrum. Defined as the ratio of noise power to carrier signal power within a 1Hz bandwidth when deviating from the carrier frequency fHz. For the single-sideband power spectral loss term, the single-sideband power spectral loss gradient can be calculated as follows:

[0103] a2. Calculate the phase power spectrum of the second phase compensation amount to obtain the estimated phase noise power spectrum;

[0104] Among these methods, mature techniques such as the average periodogram method (Bartlett method), Welch method, and Blackman-Tukey method can be used to calculate the phase power spectrum of the second phase compensation amount, thereby obtaining the estimated phase noise power spectrum. .

[0105] b2. Based on the estimated phase noise power spectrum and the prior phase noise power spectrum, calculate the single-sideband power spectrum loss value of the single-sideband power spectrum loss term;

[0106] In satellite communication systems, the single-sideband power spectrum is subject to overall performance constraints and is obtained through actual measurement using a spectrum analyzer. Therefore, it can be used as prior information for phase noise and is denoted as the prior phase noise power spectrum.

[0107] The formula for calculating the single-sideband power spectral loss value of the single-sideband power spectral loss term can be found in the relevant part of equation (2), and will not be repeated here.

[0108] c2. Differentiate the second phase compensation amount using the single-sideband power spectrum loss value to obtain the single-sideband power spectrum loss gradient of the single-sideband power spectrum loss term.

[0109] In a specific implementation, the single-sideband power spectral loss value affects the second phase compensation amount. The gradient process is as follows:

[0110] Calculate and estimate the phase noise power spectrum See calculation formula (5):

[0111] (5)

[0112] in, For FFT points, For frequency indexing.

[0113] right For the derivative, see equation (6):

[0114] (6)

[0115] Where Re denotes taking the real part, For FFT results in frequency The complex components at the location.

[0116] The chain rule is used to calculate the single-sideband power spectral loss gradient, see equation (7):

[0117] (7)

[0118] in, These are known quantities, obtained based on system specifications or measured crystal oscillator characteristics; For estimation purposes, based on the second phase compensation quantity The values ​​are obtained by performing FFT calculations. Based on these values, and by substituting them into formula (7), the single-sideband power spectral loss gradient can be calculated and used to update the generator network parameters in reverse.

[0119] In a specific implementation, the gradient of the bit error rate enhancement loss term can be calculated as follows:

[0120] a3. Obtain the padding frame data, demodulation soft information during demodulation, and re-encoding verification data after decoding from the known sequence after frame synchronization is completed;

[0121] b3. Calculate the bit error rate enhancement loss value using the padding frame data, the demodulation soft information, and the recoding check data;

[0122] The calculation formula for this process is shown in equation (8):

[0123] (8)

[0124] in, To fill in frame data or decoded re-encoded check data, To demodulate soft information, The calculation formula is shown in equation (9):

[0125] (9)

[0126] Where y is the received signal after phase compensation, i.e. , Let y be the probability that the transmitted information is 0 when the received signal is y. For the received bits. In a Gaussian white noise channel, the formula above is calculated as equation (10):

[0127] (10)

[0128] in, The variance of the white noise is the received complex signal. j represents an imaginary number, S0 is the set of constellation point symbols with a specified bit value of 0, S1 is the set of constellation point symbols with a specified bit value of 1, and s is a complex constellation point with a real part of . The imaginary part is , recorded as .

[0129] c3. Differentiate the second phase compensation amount using the bit error rate enhancement loss value to obtain the bit error rate enhancement loss gradient of the bit error rate enhancement loss term.

[0130] In a specific implementation, the process of differentiating the bit error rate enhancement loss value with respect to the second phase compensation amount is as follows:

[0131] For each bit The bit error rate enhancement loss function is related to The derivative is given in equation (11):

[0132] (11)

[0133] Assuming the received signal is after phase compensation .

[0134] The calculation of the log-likelihood ratio involves a soft decision of the sign, see equation (12), where equation (12) is a simplification of equation (10):

[0135] (12)

[0136] in, and These represent the sets of symbols where the bits are 0 and 1, respectively. The derivative is given in equation (13):

[0137] (13)

[0138] in, , .

[0139] The derivation process of formula (12) to formula (13) is as follows:

[0140] count ,but Therefore, according to the chain rule, we have the following calculation formula:

[0141]

[0142] in,

[0143] ;

[0144] Recalculate for:

[0145]

[0146] Substitute After simplification, we can obtain:

[0147] .

[0148] In summary, the formula for calculating the bit error rate enhancement loss gradient of the bit error rate enhancement loss term is given in equation (14):

[0149] (14)

[0150] In a specific implementation, the gradient of the total variation regularization loss term can be calculated as follows:

[0151] a4. Using the second phase compensation amount and the third phase compensation amount at adjacent times, calculate the total variation regularization loss value;

[0152] b4. Differentiate the second phase compensation amount using the total variation regularization loss value to obtain the total variation regularization loss gradient of the total variation regularization loss term.

[0153] The formula for calculating the gradient of the total variation regularization loss term is given in Equation (15):

[0154] (15)

[0155] The derivation process is as follows:

[0156]

[0157] in,

[0158] .

[0159] In a specific implementation, the gradient of the total variation regularization loss can be calculated as follows for the adversarial loss:

[0160] a5. Using a discriminator, calculate the confidence probability of the noise-reduced satellite communication signal during the current training based on the ideal noise-free satellite communication signal and the noise-reduced satellite communication signal during the current training.

[0161] b5. Based on the confidence probability, calculate the loss value of the adversarial loss term;

[0162] For the specific calculation formula, please refer to the calculation formula of the anti-loss term in formula (2), which will not be repeated here.

[0163] c5. Differentiate the second phase compensation amount using the loss value of the adversarial loss term to obtain the loss gradient of the adversarial loss term.

[0164] The formula for calculating the loss gradient of the adversarial loss term is given in equation (16):

[0165] (16)

[0166] in, The noise-reduced satellite communication signal for this training session, and the satellite communication signal that has been compensated using the second phase compensation amount.

[0167] In a specific implementation, the training process of the discriminator includes:

[0168] The first step is to construct a discriminant network and a discriminator loss function, wherein the discriminator loss function includes multiple second loss terms;

[0169] In a specific implementation, the input to the discrimination network is: the compensated signal constellation diagram (I / Q channels) and the ideal constellation diagram. Ideal constellation Figure 1 Generally, information is obtained through pilot signals. This invention combines satellite communication frame results and extends the information to be obtained from the frame synchronization header, padding frames, and recoding check area.

[0170] The discriminant network is structured as a convolutional neural network (CNN) for feature extraction: three convolutional layers (channel count 32→64→128). The classification layer outputs the discriminant probability through global average pooling and fully connected layers.

[0171] In a specific implementation, the formula for calculating the discriminator loss function is shown in equation (17):

[0172] (17)

[0173] The discriminator loss function is used to maximize the ability to distinguish between the real signal and the generated signal, where x is a randomly input sample noisy satellite communication signal. For an ideal noise-free satellite communication signal, it can be obtained from the frame synchronization code, pilot, padding frame, and recoded check area.

[0174] The second step is to differentiate each second loss term with respect to the second phase compensation amount obtained in the current training to obtain the second loss gradient of each second loss term.

[0175] The third step is to obtain the total loss gradient of the discriminator loss function based on each second loss gradient.

[0176] In a specific implementation, the formula for calculating the total loss gradient of the discriminator is given in equation (18):

[0177] (18)

[0178] Fourth step: Backpropagate using the total loss gradient of the discriminator loss function to update the parameters of the discriminator network until the convergence condition is met, thus obtaining the discriminator.

[0179] (3) Based on each first loss gradient, obtain the total loss gradient of the generator loss function;

[0180] In a specific implementation, the formula for calculating the total loss gradient of the generator loss function is given in equation (19):

[0181] (19)

[0182] In a specific implementation process, the above to The setup process can be followed as follows:

[0183] The first method involves setting the actual weight of the target loss term as the ratio of its expected contribution magnitude to its initial weight coefficient among the plurality of first loss terms. The target loss term may include a mean squared error loss term, a single-sideband power spectrum loss term, a code rate enhancement loss term, or a total variation regularization loss term. For example, different metrics can be assigned higher weights; for instance, satellite-to-ground communication systems prioritize bit error rate performance. This allows for assigning higher expected contribution levels. Simultaneously, to adapt to the training phase, the initial stage emphasizes the weights of the base loss (such as MSE), while later stages increase the weights of constraint terms (such as PSD and BER). For example, if initially... If the expected contribution level is 1, then .

[0184] The second method is to determine the importance of the target loss item based on the task objective, and determine the first adjustment coefficient corresponding to the importance of the target loss item based on the preset correlation between the importance and the adjustment coefficient, and adjust the current weight of the target loss item to obtain the actual weight of the target loss item.

[0185] The third method is to obtain the validation results of the validation set during each training process, and based on the preset correlation between the validation results and the adjustment coefficient, determine the second adjustment coefficient corresponding to the validation results, adjust the current weight of the target loss term, and obtain the actual weight of the target loss term.

[0186] The fourth method involves acquiring the signal characteristics of the sample signal and setting the actual weight of the target loss term based on these signal characteristics.

[0187] Specifically, the signal characteristics of a sample signal can include frequency bands, phase transition regions, etc. For example, for In this regard, higher weights are allocated to lower frequency bands. In this regard, the weights are reduced in phase transition regions (such as sign boundaries) and increased in smooth regions. Further examples will not be provided here.

[0188] The fifth method involves dynamically adjusting the actual weight of the target loss term by monitoring its gradient magnitude.

[0189] For the specific calculation formula, please refer to formula (20):

[0190] (20)

[0191] Where t represents the t-th iteration. Indicates to Find the gradient of the derivative.

[0192] (4) Backpropagate using the total loss gradient of the generator loss function to update the parameters of the generator network until the convergence condition is met, and obtain the generator.

[0193] In a specific implementation, the training process of the above generator is as follows: Figure 2 , Figure 2 This is a schematic diagram of the generator training process. (The diagram shows the process of training the generator.) Figure 2 The specific process can be as follows:

[0194] Input data preprocessing process: After the sample data is input, frequency compensation is performed to obtain the frequency-compensated signal;

[0195] The process of obtaining and compensating the phase compensation amount for the current training is as follows: extract the phase estimation features, amplitude estimation features, signal-to-noise ratio estimation features, and Doppler estimation features of the sample data to form a feature vector, input it into the generator, and after passing through the CNN layer and LSTM layer, compensate the frequency compensation signal with the phase compensation amount for the current training to obtain the noise-reduced satellite communication signal for the current training.

[0196] Mean square error loss calculation process: The signal demodulation process (including frame synchronization, decoding and output) is performed using the noise-reduced satellite communication signal of the current training. The known sequence (frame synchronization header + pilot) is extracted from the frame synchronization result. The mean square error loss is calculated in combination with the phase compensation amount of the current training.

[0197] The calculation process of single-sideband power spectrum loss is as follows: calculate the phase power spectrum of the phase compensation amount in the current training to obtain the estimated phase noise power spectrum, and calculate the single-sideband power spectrum loss by combining it with the prior phase noise power spectrum.

[0198] The calculation process of bit error rate enhancement loss is as follows: based on the frame synchronization results, fill frames are extracted and soft information is calculated. The bit error rate enhancement loss is then calculated in combination with the decoded re-encoding verification data.

[0199] The calculation process for the total variation regularization loss term is as follows: The total variation regularization loss term is calculated by combining the phase compensation amount of the current training.

[0200] The discriminator feedback process is as follows: the noise-reduced satellite communication signal of the current training is sent to the discriminator, the discriminator receives the feedback data and feeds it back to the generator, and the generator optimizes the compensation effect.

[0201] In this embodiment, adversarial learning (unsupervised) can be relied upon during pilot-free periods, while supervised loss (semi-supervised) is introduced during pilot-free periods. This reduces pilot overhead and minimizes the impact on spectral efficiency. Furthermore, various loss terms and different weight designs are employed to enhance the robustness and performance of the network.

[0202] 104. Using the first phase compensation amount, the phase noise in the frequency-compensated signal is suppressed to obtain a noise-reduced satellite communication signal.

[0203] This embodiment of the satellite communication system phase noise suppression method based on adversarial networks involves: frequency compensation of the noisy satellite communication signal to obtain a frequency-compensated signal; acquisition of the noisy signal feature vector corresponding to the noisy satellite communication signal; input of the noisy signal feature vector into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term time-series dependency data of the phase noise to obtain a first phase compensation amount; and using the first phase compensation amount to suppress the phase noise in the frequency-compensated signal to obtain a denoised satellite communication signal. This achieves the suppression of high-order QAM phase noise, thereby enabling the learning of complex time-varying noise distributions, improving dynamic tracking capabilities, and more accurately suppressing high-order QAM phase noise.

[0204] For a specific implementation process, simulation and comparison based on the above method can be found in the following:

[0205] Simulation parameters:

[0206] Modulation method: 1024QAM;

[0207] Symbol rate: 750 Msps;

[0208] Phase noise model: -80 dBc / Hz@100Hz, -125dBc / Hz@10MHz;

[0209] Residual frequency offset: 10kHz;

[0210] Pilot block length: 16 symbols;

[0211] Pilot block spacing: 10,000 symbols;

[0212] Signal-to-noise ratio: Eb / N0 = 30dB.

[0213] Simulation results:

[0214] Figure 3It is the effect of the receiver constellation diagram under different damage conditions, by Figure 3 It can be seen that phase noise and residual Doppler effects cause rotation of the received signal constellation diagram, thus leading to bit errors. Specifically, Figure 3 (1) is the original constellation diagram, (2) is the constellation diagram under phase noise, (3) is the constellation diagram under phase noise and residual frequency offset, and (4) is the constellation diagram under phase noise, residual frequency offset and white noise.

[0215] The performance of the technical solution of this invention is compared with that of existing technical solutions, and the simulation results are as follows: Figure 4 and Figure 5 , Figure 4 This refers to the phase noise suppression effects of the two schemes when there is no noise. Among them, Figure 4 (1) is the constellation diagram of the existing technical solution, with a residual phase error of 0.8° (RMS). Figure 4 (2) is the constellation diagram of the technical solution of the present invention, and its residual phase error is 0.3° (RMS).

[0216] Figure 5 The two schemes compare phase noise suppression effects when there is noise. Figure 5 Figure (1) shows the constellation diagram of the existing technical solution, with a residual phase error of 1.15° (RMS) and a bit error rate of 1.2e-3. Figure 5 (2) is the constellation diagram of the technical solution of the present invention, with a residual phase error of 0.88° (RMS) and a bit error rate of 1.8e-6.

[0217] Depend on Figures 4 to 5 It can be seen that, compared with the prior art, the technical solution of the present invention reduces the phase error by 62.5% after suppressing phase noise when there is no noise; and improves the bit error rate by three orders of magnitude when there is noise.

[0218] Based on the same general inventive concept, this invention also protects a dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks. The dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks provided by this invention will be described below. The dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks described below can be referred to in correspondence with the phase noise suppression method for a satellite communication system based on adversarial networks described above.

[0219] Figure 6 This is a schematic diagram of the dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the dynamic phase noise suppression system for a high-order satellite communication system based on adversarial networks in this embodiment includes a first compensation module 61, an acquisition module 62, a second compensation module 63, and a suppression module 64.

[0220] The first compensation module 61 is used to perform frequency compensation on the noisy satellite communication signal to obtain a frequency-compensated signal.

[0221] The acquisition module 62 is used to acquire the noisy signal feature vector corresponding to the noisy satellite communication signal;

[0222] The second compensation module 63 is used to input the noisy signal feature vector into a generator pre-trained based on an adversarial network for calculation to obtain the first phase compensation amount;

[0223] The suppression module 64 is used to suppress the phase noise in the frequency-compensated signal using the first phase compensation amount to obtain a noise-reduced satellite communication signal.

[0224] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The high-order satellite communication system dynamic phase noise suppression system based on adversarial networks may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740. The processor 710, communication interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute the satellite communication system phase noise suppression method based on adversarial networks.

[0225] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0226] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the phase noise suppression method for satellite communication systems based on adversarial networks provided by the above methods.

[0227] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the phase noise suppression method for satellite communication systems based on adversarial networks provided by the above methods.

[0228] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0229] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0230] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A phase noise suppression method for a satellite communication system based on adversarial networks, characterized in that, include: Frequency compensation is performed on the noisy satellite communication signal to obtain a frequency-compensated signal; Obtain the noisy signal feature vector corresponding to the noisy satellite communication signal; The noisy signal feature vector is input into a generator pre-trained based on an adversarial network to extract local fluctuation data and long-term temporal dependence data of phase noise, thereby obtaining the first phase compensation amount. Using the first phase compensation amount, the phase noise in the frequency-compensated signal is suppressed to obtain a noise-reduced satellite communication signal; The training process of the generator includes: Construct a generator network and a generator loss function, wherein the generator loss function includes multiple first loss terms; The first loss gradient of each first loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each first loss term. Based on each first loss gradient, the total loss gradient of the generator loss function is obtained; Backpropagation is performed using the total loss gradient of the generator loss function to update the parameters of the generator network until the convergence condition is met, thus obtaining the generator. The generator loss function includes an adversarial loss term, the gradient of which is obtained based on the discriminator; the training process of the discriminator includes: Construct a discriminant network and a discriminator loss function, wherein the discriminator loss function includes multiple second loss terms; The second loss gradient of each second loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each second loss term. Based on each second loss gradient, the total loss gradient of the discriminator loss function is obtained; Backpropagation is performed using the total loss gradient of the discriminator loss function to update the parameters of the discriminator network until the convergence condition is met, thus obtaining the discriminator.

2. The phase noise suppression method for satellite communication systems based on adversarial networks according to claim 1, characterized in that, The generator loss function includes a mean squared error loss term; Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including: The maximum likelihood algorithm is used to calculate the actual required compensation amount based on the received data, transmitted data, frame synchronization length, and pilot block length in the known sequence after frame synchronization is completed. Based on the actual demand compensation amount and the second phase compensation amount, the mean square error of the mean square error loss term is calculated; The mean square error loss gradient of the mean square error loss term is obtained by differentiating the second phase compensation amount using the mean square error.

3. The phase noise suppression method for satellite communication systems based on adversarial networks according to claim 1, characterized in that, The generator loss function includes a single-sideband power spectrum loss term; Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including: The phase power spectrum of the second phase compensation is calculated to obtain the estimated phase noise power spectrum; Based on the estimated phase noise power spectrum and the prior phase noise power spectrum, the single-sideband power spectrum loss value of the single-sideband power spectrum loss term is calculated. The single-sideband power spectrum loss gradient of the single-sideband power spectrum loss term is obtained by differentiating the second phase compensation amount with the single-sideband power spectrum loss value.

4. The phase noise suppression method for satellite communication systems based on adversarial networks according to claim 1, characterized in that, The generator loss function includes a bit error rate enhancement loss term; Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including: Acquire padding frame data, demodulation soft information during demodulation, and re-encoding verification data after decoding from the known sequence after frame synchronization is completed; The bit error rate enhancement loss value is calculated using the padding frame data, the demodulation soft information, and the recoding check data. The bit error rate enhancement loss gradient of the bit error rate enhancement loss term is obtained by differentiating the second phase compensation amount with the bit error rate enhancement loss value.

5. The phase noise suppression method for satellite communication systems based on adversarial networks according to claim 1, characterized in that, The generator loss function includes a total variation regularization loss term; Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including: The total variation canonical loss value is calculated using the second phase compensation amount and the third phase compensation amount at adjacent time points. The total variation regularization loss gradient of the total variation regularization loss term is obtained by differentiating the second phase compensation amount with the total variation regularization loss value.

6. The phase noise suppression method for satellite communication systems based on adversarial networks according to claim 1, characterized in that, The generator loss function includes an adversarial loss term; Differentiate each first loss term with respect to the second phase compensation obtained in the current training iteration to obtain the first loss gradient of each first loss term, including: Using a discriminator, the confidence probability of the noise-reduced satellite communication signal during the current training is calculated based on the ideal noise-free satellite communication signal and the noise-reduced satellite communication signal during the current training. Based on the confidence probability, calculate the loss value of the adversarial loss term; The loss gradient of the adversarial loss term is obtained by differentiating the second phase compensation amount with the loss value of the adversarial loss term.

7. The phase noise suppression method for a satellite communication system based on adversarial networks according to any one of claims 1 to 6, characterized in that, Also includes: The ratio of the expected contribution magnitude of the target loss term among the plurality of first loss terms to the initial weight coefficient of the target loss term is set as the actual weight of the target loss term; or Based on the task objective, the importance of the target loss item is determined, and based on the preset correlation between the importance and the adjustment coefficient, a first adjustment coefficient corresponding to the importance of the target loss item is determined, and the current weight of the target loss item is adjusted to obtain the actual weight of the target loss item. or Obtain the validation results of the validation set during each training process, and based on the preset correlation between the validation results and the adjustment coefficient, determine the second adjustment coefficient corresponding to the validation results, adjust the current weight of the target loss term, and obtain the actual weight of the target loss term; or Obtain the signal characteristics of the sample signal, and set the actual weight of the target loss term based on the signal characteristics; or The actual weight of the target loss term is dynamically adjusted by monitoring the gradient magnitude of the target loss term.

8. A dynamic phase noise suppression device for a high-order satellite communication system based on adversarial networks, characterized in that, include: The first compensation module is used to perform frequency compensation on noisy satellite communication signals to obtain a frequency-compensated signal. The acquisition module is used to acquire the noisy signal feature vector corresponding to the noisy satellite communication signal; The second compensation module is used to input the noisy signal feature vector into a generator pre-trained based on an adversarial network, extract local fluctuation data and long-term time-series dependency data of phase noise, and obtain the first phase compensation amount. The suppression module is used to suppress the phase noise in the frequency-compensated signal using the first phase compensation amount, so as to obtain a noise-reduced satellite communication signal. The training process of the generator includes: Construct a generator network and a generator loss function, wherein the generator loss function includes multiple first loss terms; The first loss gradient of each first loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each first loss term. Based on each first loss gradient, the total loss gradient of the generator loss function is obtained; Backpropagation is performed using the total loss gradient of the generator loss function to update the parameters of the generator network until the convergence condition is met, thus obtaining the generator. The generator loss function includes an adversarial loss term, the gradient of which is obtained based on the discriminator; the training process of the discriminator includes: Construct a discriminant network and a discriminator loss function, wherein the discriminator loss function includes multiple second loss terms; The second loss gradient of each second loss term is obtained by taking the derivative of the second phase compensation amount obtained in the current training for each second loss term. Based on each second loss gradient, the total loss gradient of the discriminator loss function is obtained; Backpropagation is performed using the total loss gradient of the discriminator loss function to update the parameters of the discriminator network until the convergence condition is met, thus obtaining the discriminator.

Images