Method and device for recovering time-domain nonlinear distortion of OFDM signal of satellite communication

Abstract

This invention discloses a method and system for recovering time-domain nonlinear distortion of OFDM signals in satellite communication, mainly addressing the problem of difficulty in recovering nonlinear distortion of OFDM signals in existing technologies. The solution includes: acquiring a dataset for recovering time-domain nonlinear distortion of OFDM signals in satellite communication; constructing a deep learning network model consisting of an optimized multi-layer one-dimensional convolutional neural network and a bidirectional long short-term memory network module; inputting the acquired dataset into the deep learning network model for training, verification, and testing to recover the time-domain distorted signal; and performing channel equalization and demodulation on the recovered time-domain signal to complete the demodulation from modulation symbols to bit information, achieving effective data recovery and transmission. This invention exhibits good mean square error and bit error rate performance, reduces the complexity of the network model, effectively recovers signal distortion caused by nonlinear distortion, and improves the accuracy and efficiency of signal recovery. It can be used for the recovery of nonlinear distorted signals in satellite communication systems.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and specifically relates to a method and apparatus for restoring OFDM signal nonlinear distortion. It can be used to efficiently restore OFDM signal nonlinear distortion caused by nonlinear power amplifiers, multipath effects and noise interference in satellite communication systems, thereby improving the signal recovery quality and the overall performance of the system. Background Technology

[0002] In satellite communication systems, Orthogonal Frequency Division Multiplexing (OFDM) technology has become one of the important modulation techniques for modern satellite communication due to its strong resistance to multipath fading and high spectral efficiency. However, due to the special nature of the satellite communication environment, signals are affected by many factors during transmission, such as nonlinear distortion introduced by nonlinear power amplifiers, multipath effects, and noise interference, resulting in severe distortion of the received OFDM signal. If this distortion is not effectively recovered, it will seriously affect the performance of the communication system, especially under high-order modulation and low signal-to-noise ratio (SNR) conditions, where the presence of distortion will greatly reduce the accuracy of demodulation, leading to a significant increase in the bit error rate (BER).

[0003] To recover distorted signals, traditional methods typically rely on model-based channel estimation and equalization algorithms, such as the Minimum Mean Square Error (MMSE) algorithm and the Zero Forcing (ZF) algorithm. However, model-based methods usually require pre-assuming channel and distortion models, and their performance can be significantly limited in complex communication environments.

[0004] With the development of deep learning technology, data-driven deep learning methods have gradually demonstrated their powerful capabilities in complex nonlinear problems. Against this backdrop, deep learning-based OFDM signal recovery methods have emerged. By learning from a large amount of distorted signal data, deep learning models can effectively capture the complex nonlinear characteristics of the channel and achieve accurate recovery of time-domain signals without relying on a definite channel model.

[0005] Reference 1, “Z. Huang, D. He and Z. Wang, Two-Stage LMMSE / DNN Receiver for High-Order Modulation,” in IEEE Communications Letters, vol. 27, no. 8, pp. 2068-2072, Aug. 2023,” discloses a two-stage LMMSE / DNN receiver to improve demodulation performance of OFDM signals in the presence of multipath effects and nonlinear distortion. This receiver consists of two parts: an LMMSE equalizer and a DNN. Specifically, the received signal is first processed by an LMMSE equalizer to reduce multipath effects, and then further processed by a DNN to eliminate nonlinear distortion. However, the designed network model is relatively simple, making it difficult to adapt to complex nonlinear systems.

[0006] Patent application CN201310703252.6 discloses an efficient joint method for nonlinear distortion recovery and channel estimation in OFDM ultra-wideband systems. This method performs two amplitude limiting operations at the transmitter to reduce the signal amplitude below a preset threshold, and at the receiver, performs nonlinear distortion estimation and compensation based on compressed sensing-based sparse channel estimation and equalization. While this method can suppress signal nonlinear distortion caused by the peak-to-average power ratio (PAPR) of OFDM ultra-wideband systems and reduce channel estimation performance, it suffers from high computational complexity and does not consider the nonlinear distortion effect of the power amplifier, thus limiting the system's nonlinear effects.

[0007] Patent document CN201610512985.5 discloses a joint estimation algorithm for OFDM compressed sensing channel and nonlinear distortion. This algorithm simultaneously estimates the nonlinear distortion caused by the channel and power amplification in the receiver to avoid the complex peak-to-average power ratio (PAPR) algorithm at the transmitter. However, because it is highly dependent on channel characteristics and requires the signal to have strong sparsity, it is difficult to handle signals with complex nonlinearity and low sparsity.

[0008] Furthermore, some existing methods for channel estimation and signal recovery using neural networks require training with large simulation datasets to allow the network to gradually learn the nonlinear characteristics and distortion patterns of the channel. However, when applied to satellite communications, especially for signal distortion recovery tasks caused by power amplifiers, these methods still face problems such as insufficient datasets, high network complexity, and overfitting. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of the prior art by proposing a method and apparatus for restoring the time-domain nonlinear distortion of OFDM signals in satellite communication, so as to restore OFDM signal distortion caused by factors such as nonlinear power amplifiers, multipath effects and noise interference in satellite communication systems, thereby improving signal quality and the overall performance of satellite communication systems.

[0010] To achieve the above objectives, the present invention provides a method for recovering time-domain nonlinear distortion of OFDM signals in satellite communication, comprising:

[0011] Acquire a dataset of time-domain nonlinear distortion recovery for OFDM signals from satellite communications;

[0012] Construct a deep learning network model consisting of N layers of one-dimensional convolutional neural networks and L bidirectional long short-term memory networks;

[0013] The acquired time-domain nonlinear distortion recovery dataset is input into the constructed deep learning network model for training to recover the time-domain distortion signal.

[0014] For the recovered time-domain signal Channel equalization and demodulation are performed to obtain the final bit stream b.

[0015] Furthermore, the construction of the deep learning network model includes the following implementation:

[0016] a) Construct an N (N≥2) layer one-dimensional convolutional neural network, where the first convolutional layer receives a single-channel input signal y(n) and generates a feature map with m channels. The middle convolutional layers maintain m channels, with a kernel size of i×i and padding of j. ReLU activation function is applied after each convolutional layer to normalize the signal and introduce non-linear characteristics of the network. The last convolutional layer converts the feature map of m channels back to a single channel and outputs the initially recovered signal sequence y(t).

[0017] b) Input the initially recovered signal sequence y(t) into the existing L bidirectional long short-term memory network modules, and perform residual skip connections between the first and last convolutional layers to obtain the enhanced signal sequence y(m);

[0018] c) Input the enhanced signal sequence y(m) into the fully connected layer, and output the recovered signal sequence after passing through the fully connected layer. This means completing the construction of a deep learning network model.

[0019] Furthermore, the channel equalization and demodulation of the recovered signal are implemented as follows:

[0020] (a) The recovered signal By removing the cyclic prefix (CP) and performing a Fast Fourier Transform (FFT) operation, the frequency domain signal Y at the receiving end is obtained. K ;

[0021] (b) Extract the frequency domain signal Y K pilot symbols in By performing minimum linear mean square error (LMMSE) channel estimation, the channel response matrix at the pilot symbol is obtained.

[0022] (c) Channel response matrix at pilot symbols Perform linear interpolation to obtain the channel response matrix H of the entire signal. K ;

[0023] (d) For the channel response matrix H K The result X after equilibrium is obtained by performing minimum mean square error MMSE equilibrium. K ;

[0024] (e) Perform a sign decision on the equalized result X according to the constellation diagram rules, and demap the decision result to the corresponding bit data, i.e. demodulate the bit data b.

[0025] To achieve the above objectives, the present invention provides a satellite communication OFDM signal time-domain nonlinear distortion recovery device, comprising:

[0026] The digital modulation module is used to digitally modulate the bit stream and map it into frequency domain complex symbols;

[0027] The OFDM modulation module is used to convert frequency domain complex symbols into time domain signals through inverse fast Fourier transform (IFFT), while adding a cyclic prefix (CP) to reduce multipath interference.

[0028] The power amplifier module is used to amplify the power of the OFDM modulated signal to ensure the transmission power of the signal in the satellite communication link;

[0029] The upconversion and radio frequency module is used to convert the amplified baseband signal into a radio frequency signal and transmit it to the satellite through the antenna to realize wireless signal transmission;

[0030] The receiving radio frequency and down-conversion module is used to receive radio frequency signals from satellites and down-convert them back to baseband signals in preparation for subsequent signal processing.

[0031] The time-domain restoration network module is used to perform preliminary nonlinear distortion recovery processing on the received satellite signals;

[0032] The OFDM demodulation module is used to perform Fast Fourier Transform (FFT) on the initially recovered signal to remove pilot symbols and cyclic prefixes.

[0033] The channel estimation and equalization module is used to estimate channel characteristics and perform equalization processing to compensate for fading and distortion in the channel and restore the original transmitted signal.

[0034] The digital demodulation module is used to demodulate the equalized signal into a bit stream, thus completing the data demodulation.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] Firstly, by combining the time-domain characteristics of OFDM signals in satellite communication, this invention designs a deep learning model composed of an optimized one-dimensional convolutional neural network module and a bidirectional long short-term memory network module, which can effectively capture the time-domain characteristics of the channel and power amplifier, and improve the accuracy and efficiency of signal distortion recovery.

[0037] Secondly, the present invention only performs distortion recovery at the receiving end, without requiring any processing at the transmitting end. Therefore, it greatly reduces computational overhead and can complete the recovery of distorted signals in a shorter time, thereby improving the real-time performance and processing efficiency of the system.

[0038] Third, this invention utilizes a neural network to recover nonlinear distortion at the receiving end, followed by channel estimation and equalization to further recover the distortion caused by the channel. This avoids using a large amount of data for network training and reduces the probability of overfitting. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the implementation of the OFDM signal time-domain nonlinear distortion recovery method for satellite communication according to the present invention.

[0040] Figure 2 This is a network structure diagram constructed in an embodiment of the present invention;

[0041] Figure 3 This is a block diagram of the OFDM signal time-domain nonlinear distortion recovery device for satellite communication according to the present invention;

[0042] Figure 4 This is a simulation comparison of the mean square error (MSE) of the present invention and existing nonlinear distortion recovery methods;

[0043] Figure 5 This is a simulation comparison chart of the bit error rate (BER) of the present invention and existing nonlinear distortion recovery methods. Detailed Implementation

[0044] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0045] The key technology of this invention lies in constructing a one-dimensional convolutional bidirectional long short memory network 1D-CNN-BiLSTM and applying it to existing satellite communication OFDM systems to achieve the recovery of time-domain nonlinear distortion signals.

[0046] Example 1: Method for restoring time-domain nonlinear distortion of OFDM signals in satellite communication.

[0047] Reference Figure 1 The implementation steps of this example include the following:

[0048] Step 1: Obtain the time-domain nonlinear distortion recovery dataset of OFDM signals from satellite communication.

[0049] This step is implemented based on a satellite communication OFDM system, which includes transmitter digital modulation, transmitter OFDM modulation, power amplifier, satellite channel, receiver OFDM demodulation, and receiver digital demodulation. The specific steps are as follows:

[0050] 1.1) At the transmitting end of the satellite communication system, the random function randi() is first used to generate a generator with a length of... binary bit stream Each a i ∈{0,1}, Q is the number of symbols, P is the number of subcarriers, and M is the modulation order; this example uses 16QAM modulation, i.e., M=16, the number of symbols Q is set to 60, and the subcarrier format P is set to 512. The generated binary bit stream a is divided into several groups in sequence, each group containing Each group of 4 bits is mapped to a point on a two-dimensional constellation diagram according to the 16QAM modulation scheme. The real and imaginary parts of each 16QAM symbol correspond to a coordinate point in the constellation diagram. The coordinates of this point are determined by a predefined mapping rule. The mapped 16QAM symbols are arranged in order to generate the final modulated signal.

[0051] 1.2) Perform an inverse fast Fourier transform (IFFT) on each group of Q symbols to convert the frequency domain signal into a time domain signal x(s):

[0052]

[0053] Among them, X k Let P be the 16QAM symbol on the k-th subcarrier, where P is the number of subcarriers, which can also be represented as the number of points in the IFFT.

[0054] 1.3) In order to eliminate inter-symbol interference (ISI) caused by multipath effect, a vertical pilot is inserted before the time domain signal x(s), a cyclic prefix (CP) is added, and the time domain signal after adding the cyclic prefix is ​​converted from parallel to serial to generate a continuous signal x(n), which is used as the label of the dataset.

[0055] 1.4) Calculate the nonlinear characteristics of the power amplifier to obtain the correspondence between the magnitude of the output signal and the magnitude of the input signal G(|x(n)|) and the phase difference relationship between the output signal and the input signal φ(|x(n)|):

[0056]

[0057] Where, α ρ β ρ α θ β θ To describe the four different parameters of the system, in this example, the values ​​are, but are not limited to, α. ρ =2,β ρ =1, β θ =1;

[0058] 1.5) The parallel-to-serial conversion signal x(n) is amplified by a nonlinear power amplifier HPA to output the amplified signal z(n):

[0059] z(n) = G(|x(n)|)e jθ(x(n)+φ(|x(n)|))

[0060] 1.6) The amplified signal z(n) is transmitted through the satellite channel h with Gaussian white noise w(n) added to obtain the received signal y(n), which is used as a sample of the dataset:

[0061] y(n) = z(n) * h + w(n)

[0062] This embodiment uses a Lutz channel, but is not limited to this channel;

[0063] 1.7) Generate N sets of samples y(n) and labels x(n) by looping N times, and use the obtained samples and labels to form a dataset.

[0064] Step 2: Preprocess the dataset consisting of samples y(n) and labels x(n) by dividing it into training set, validation set and test set.

[0065] 2.1) Transform the dimensions of the samples y(n) and labels x(n) in the dataset, extract their real and imaginary parts, and flatten them into a one-dimensional sequence, so that each signal sample becomes a long sequence composed of real and imaginary parts to adapt to the input rules of the network.

[0066] 2.2) Normalization is used to adjust the numerical range of the real and imaginary parts to the range of [0,1], making the data more stable and helping the network converge faster;

[0067] 2.2) Divide the normalized dataset into training, validation, and test sets in a ratio of 7:1.5:1.5;

[0068] The training set is used to train the neural network model and is the basic data for model learning;

[0069] The validation set is used to evaluate the model's performance during training, help tune hyperparameters, and prevent overfitting.

[0070] The test set is used to evaluate the final performance of the model after training and to simulate the model's performance in real-world applications.

[0071] Step 3: Construct a deep learning network model.

[0072] Existing fully connected deep neural networks (DNNs) are prone to overfitting and getting stuck in local optima due to their large number of parameters. Furthermore, they cannot analyze the overall logical sequence of input information or the complex temporal relationships between information, thus failing to model changes in time series. Denoising convolutional neural networks (DnCNNs), through multi-layer convolution operations, can effectively extract local features from time-domain signals. This local feature extraction capability helps the model better identify and suppress noise and interference caused by nonlinear distortion. Bidirectional long short-term memory (BiLSTM) networks, through their bidirectional forward and backward processing capabilities, capture the complex dependencies between signal points in the time-domain signal, thereby significantly improving the accuracy and stability of nonlinear distortion signal recovery. Therefore, this example constructs a network model composed of a denoising convolutional network (DnCNN) and a bidirectional long short-term memory network (BiLSTM).

[0073] refer to Figure 2 The specific implementation method of this step is as follows:

[0074] 3.1) The existing denoising convolutional network DnCNN model is optimized to reduce computational complexity, that is, the original 17-layer DnCNN is simplified into a 7-layer one-dimensional convolutional neural network 1D-CNN, where:

[0075] The first convolutional layer receives a single-channel input signal y(n) and generates a feature map with 32 channels;

[0076] The middle 5-6 convolutional layers maintain 32 channels, with each kernel size of 3×3 and padding set to 1 to keep the feature map size constant. ReLU activation and batch normalization are performed sequentially after each convolutional layer. ReLU activation accelerates neural network convergence and alleviates overfitting. Batch normalization involves inversely normalizing the preprocessed data based on the mean and variance to reduce internal covariate shifts and prevent slow model fitting.

[0077] The final layer, the 7th convolutional layer, converts the feature maps of the 32 channels back to a single channel, and outputs the initially recovered signal sequence y(t).

[0078] 3.2) The initially recovered signal sequence y(t) is input into an existing bidirectional long short-term memory network module, and a residual skip connection is made between the first and last convolutional layers to obtain the enhanced signal sequence y(m). The residual connection is added to directly transmit the low-frequency information of the initial input signal to the output layer, improve training efficiency and model performance, and alleviate the problem of gradient vanishing or exploding.

[0079] 3.3) Select a fully connected layer containing 128 neurons, input the enhanced signal sequence y(m) into the fully connected layer for linear transformation, nonlinear activation and feature extraction, and output the recovered signal sequence. This sequence is a long string of single-channel signals, which has the same dimension as the input time-domain signal, thus completing the construction of the deep learning network model.

[0080] Step 4: Train and validate the constructed deep learning network model.

[0081] 4.1) Set the minimum batch size, learning rate η, number of training epochs, and total number of samples N;

[0082] 4.2) Initialize the parameters in the deep learning network model to give it an initial state in order to avoid gradient vanishing or gradient exploding in the early stages of training;

[0083] 4.3) Input the divided training set into the network for forward propagation:

[0084] Forward propagation is the first stage in the computation process of a neural network. In this stage, the input data is first linearly transformed by the weight matrix in each convolutional layer; then nonlinearly transformed by the ReLU activation function; and finally normalized by batch normalization to generate the input for the next layer. This process is repeated until the last layer of the network, and finally the forward propagation result is output.

[0085] 4.4) After each forward propagation, the mean squared error (MSE) is used as the loss function to calculate the error between the network output and the true label:

[0086]

[0087] Where N is the total number of samples, and x(n) is the label. It is the recovered signal sequence;

[0088] 4.5) Calculate the gradient ▽ of the loss function with respect to the network parameters based on the error. θJ(θ):

[0089] 4.6) Use the backpropagation algorithm to propagate the gradient from the output layer back to the input layer to update the gradient, thereby reducing the output error and obtaining the current gradient:

[0090] θ t+1 =θ t -η▽ θ J(θ)

[0091] Where, θ t+1 θ represents the currently updated parameter. t This represents the current parameter, where η is the learning rate, ▽ θ J(θ) is the gradient of the loss function with respect to the parameter θ, which indicates in which direction the parameter should be updated to reduce the loss;

[0092] 4.7) The Adam optimizer is used to update the network parameters step by step based on the calculated gradient and learning rate. At the same time, the existing ReduceLROnPlateau learning rate optimizer is used to dynamically adjust the learning rate until the set training epoch is reached to complete the training.

[0093] 4.8) Input the pre-defined validation set into the trained deep learning network model for forward propagation, repeat step 4.4), calculate the MSE loss function value for each epoch, until the set training epoch is reached, and determine whether the MSE loss function value for each epoch is close to the loss value obtained by the network during training:

[0094] If they are similar, it indicates that the constructed network model has good robustness and generalization ability. In this case, the network model parameters of the last round are saved as the optimal network model parameters.

[0095] If the two differ significantly, it is necessary to adjust parameters such as the learning rate and loss function during the training process and retrain and validate.

[0096] In this example, the total number of samples and labels is set to 2000. Mini-batch gradient descent is used during training, with 32 mini-batches per epoch, for a total of 100 epochs. The initial learning rate is set to 0.001, and each time the learning rate decreases, it is updated by 0.5 times the current learning rate.

[0097] Step 5: Recover the signal sequence using the validated network model.

[0098] 5.1) Input the test set into the validated network model, load its optimal network model parameters, perform one forward propagation, and calculate the value of the MSE loss function;

[0099] 5.2) Determine if the value of the MSE loss function is close to the loss value obtained in the last round of network training and validation:

[0100] If the samples are similar, it indicates that the constructed network model has good robustness and generalization ability. In this case, a new set of distorted samples is generated according to step 1 and input into the network model to obtain the recovered signal sequence.

[0101] If the two differ significantly, it is necessary to adjust parameters such as the learning rate and loss function during the training process and retrain and validate.

[0102] Step 6: Process the recovered signal sequence Perform channel estimation, equalization, and demodulation.

[0103] 6.1) For the recovered signal sequence By removing the cyclic prefix (CP) and performing a Fast Fourier Transform (FFT) operation, the frequency domain signal Y at the receiving end is obtained. K :

[0104] Y K =X K H K +W K

[0105] Among them, X K H is the frequency domain signal transmitted by the transmitting end. K For the channel frequency domain response, W K This is frequency domain noise.

[0106] 6.2) Extract the frequency domain signal Y from the receiving end. K pilot signal in

[0107] 6.2.1) In the frequency domain signal Y K In this process, the location of the pilot signal is found according to the 16QAM modulation scheme and signal transmission protocol. This pilot signal is a known signal and is usually used for synchronization and channel estimation.

[0108] 6.2.2) Based on the found pilot signal positions, perform indexing and slicing operations to extract the pilot signal from the receiver. 6.3) Pilot signal at the receiving end The channel response matrix is ​​obtained by performing minimum linear mean square error (LMMSE) channel estimation.

[0109] 6.3.1) Assume the channel autocorrelation matrix R HH It is a diagonal matrix of the real channel H, used to avoid complex matrix inversion operations, thereby reducing computational complexity;

[0110] 6.3.2) For the pilot signal at the receiving end Perform least squares estimation to obtain the initial estimate.

[0111]

[0112] in This is the pilot signal for the transmitting end;

[0113] 6.3.3) Set the channel estimation constant β for 16QAM modulation. In this example, the value of β is 17 / 9, but not limited to 17 / 9.

[0114] 6.3.4) Based on the channel autocorrelation matrix R HH and the results of the initial estimate Calculate the channel response matrix of the pilot signal at the receiving end.

[0115]

[0116] Where SNR is the signal-to-noise ratio of the signal, and I is the identity matrix.

[0117] 6.4) Channel response matrix for the received pilot signal Perform linear interpolation to obtain the channel response matrix H of the entire signal. K :

[0118] 6.4.1) Determine the channel response matrix of the pilot signal at the receiving end. The location of the missing data points;

[0119] 6.4.2) Find the coordinates of the adjacent data points of the missing data point, and use the linear interpolation formula to calculate the y-value of the missing data point at x:

[0120] y = y1 + (x - x1) * (y2 - y1) / (x2 - x1)

[0121] Where (x,y) is the coordinate of a missing data point, and (x1,y1) and (x2,y2) are the coordinates of its neighboring points;

[0122] 6.4.3) Fill in the coordinates of all the missing data points into the corresponding positions to obtain the channel response matrix H of the entire signal. K .

[0123] 6.5) Regarding the channel response matrix H K The result X after equilibrium is obtained by performing minimum mean square error MMSE equilibrium. K :

[0124] 6.5.1) Using the average power of the transmitted signal, the average power of the noise, and the frequency domain channel response matrix H K Determine the MMSE equalizer coefficient C MMSE :

[0125]

[0126] in, The average power of the noise, H represents the average power of the transmitted signal. K * H represents K The conjugate matrix of , where I is the identity matrix;

[0127] 6.5.2) Using MMSE equalizer coefficient C MMSE Calculate the equilibrium result X K :

[0128] X K =C MMSE Y K

[0129] Among them, Y K This is the frequency domain signal received at the receiver.

[0130] 6.6) Equilibrium result X K According to the constellation diagram rules, a sign decision is made, and the decision result is demapped into the corresponding bit data, i.e., the demodulated bit data b is obtained:

[0131] 6.6.1) The equilibrium result X K Each signal in the diagram is mapped onto a constellation diagram corresponding to its modulation scheme, and the Euclidean distance d between the signal and each point in the constellation diagram is calculated. ij :

[0132]

[0133] Among them, (x i ,y i ) represents the coordinates of each signal in the equalization result X, (x j ,y j () represents the coordinates of each point in the constellation chart;

[0134] 6.6.2) Based on the distance calculation results, the constellation point with the closest distance is selected as the decision result. The purpose is to minimize the symbol error rate and ensure that the transmitted symbols are recovered as accurately as possible in noisy and interference environments.

[0135] 6.6.3) Map the decision result points to the original bit stream b, complete the demodulation process from modulation symbols to bit information, and thus realize the effective recovery and transmission of data.

[0136] refer to Figure 3 The present invention provides a time-domain nonlinear distortion recovery device for OFDM signals in satellite communication, comprising a digital modulation module 1, an OFDM modulation module 2, a power amplification module 3, an up-conversion and radio frequency module 4, a receiving radio frequency and down-conversion module 5, a time-domain restoration network module 6, an OFDM demodulation module 7, a channel estimation and equalization module 8, and a digital demodulation module 9.

[0137] The working principle of the entire device is as follows:

[0138] Digital modulation module 1 performs 16QAM modulation on the input bitstream, mapping the binary data to specific points on the complex plane to form a frequency domain symbol, which is then transmitted to OFDM modulation module 2. OFDM modulation module 2 performs Inverse Speed ​​Fourier Transform (IFFT) on the frequency domain symbol, converts it into a time domain signal, adds pilot signals and a cyclic prefix (CP), and outputs it to power amplifier module 3. Power amplifier module 3 amplifies the OFDM modulated signal to ensure sufficient transmission power in the satellite communication link and transmits the amplified signal to upconversion and RF module 4. Upconversion and RF module 4 upconverts the power-amplified OFDM signal to the RF band, and the RF signal is transmitted to the satellite through an antenna to achieve wireless signal transmission.

[0139] Down-conversion module 5 receives the RF signal from the satellite and down-converts it back to baseband signal, then transmits it to time-domain restoration network module 6. Time-domain restoration network module 6 performs preliminary nonlinear distortion recovery processing on the received satellite signal using a neural network. The recovered signal is then transmitted to the subsequent OFDM demodulation module 7. OFDM demodulation module 7 first performs a Fast Fourier Transform (FFT) on the signal initially recovered from the time-domain multiplexing network to restore the frequency domain symbols. Simultaneously, it removes the cyclic prefix (CP) and transmits the signal to channel estimation and equalization module 8. Channel estimation and equalization module 8 performs channel estimation on the signal after removing the pilot and CP to obtain channel characteristic parameters, such as attenuation and delay, and uses these parameters for equalization processing to compensate for fading and distortion in the channel, ensuring further accurate signal recovery. The equalized signal is then transmitted to digital demodulation module 9. Digital demodulation module 9 uses a 16QAM demapping rule corresponding to the modulation mapping rule to restore the frequency domain symbols to the original binary bit data, ensuring accurate acquisition of the bit information transmitted by the transmitter.

[0140] The effects of this invention can be further illustrated by the following simulation results.

[0141] I. Simulation Conditions

[0142] The modulation scheme is 16QAM, the number of symbols is 60, the number of pilots is 3, the CP length is 32, the number of FFT points is 512, the number of training samples is 2000, the initial learning rate is 0.001, the number of training rounds is 100, and the signal-to-noise ratio range is 0-20dB.

[0143] The simulation platforms are MATLAB R2023b and Python 3.7, pytorch-gpu version. The required libraries are: os, numpy, sklearn, matplotlib, and torch.

[0144] The training process was based on a Python 3.7 environment and deployed on an NVIDIA RTX 3080 GPU using the PyTorch deep learning framework.

[0145] II. Simulation Content

[0146] Simulation 1: Under the above conditions, the present invention and existing TSLD, LMMSE, and DnCN methods were used to perform nonlinear recovery processing on OFDM time-domain distortion signals, and their respective MSE errors were calculated. The results are as follows: Figure 4 .

[0147] from Figure 4 As can be seen, within the signal-to-noise ratio range of 0-20dB, the present invention has the lowest MSE error compared to other existing methods, demonstrating the superior performance of the present invention in terms of MSE error.

[0148] Simulation 2: Under the above conditions, the present invention and existing TSLD, LMMSE, and DnCN methods were used to perform nonlinear recovery processing on OFDM time-domain distorted signals, and their bit error rates were calculated. The results are as follows: Figure 5 .

[0149] from Figure 5 As can be seen, when the signal-to-noise ratio (SNR) is low, the bit error rate (BER) performance of this invention is similar to that of several existing methods. When the SNR is high, this invention exhibits a significant advantage in BER performance compared to other methods. Specifically, at an SNR of 10 dB, the BER of this invention is approximately 0.0527, while the BER of the existing DnCNN method is 0.0628, the BER of the existing TSLD method is approximately 0.0636, and the BER of the existing LMMSE method is approximately 0.0724, demonstrating that this invention has the best BER performance. When the SNR is 20 dB, the BER of this invention is approximately 0.0201, while the BER of the existing DnCNN method is approximately 0.0232, the BER of the existing TSLD method is approximately 0.0362, and the BER of the existing LMMSE method is approximately 0.0605, demonstrating that this invention offers a greater improvement in BER performance compared to existing methods at high SNR.

[0150] The simulation results show that the present invention has good mean square error and bit error rate performance, can reduce the complexity of the original model implementation, effectively recover signal distortion caused by nonlinear distortion, and improve the accuracy and efficiency of signal recovery.

[0151] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0152] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

Claims

1. A method for recovering time-domain nonlinear distortion of OFDM signals in satellite communication, characterized in that, include: Acquire a dataset of time-domain nonlinear distortion of OFDM signals from satellite communications; A deep learning network model consisting of N layers of one-dimensional convolutional neural networks and L bidirectional long short-term memory networks is constructed, and its implementation includes the following: 3a) Construction One-layer one-dimensional convolutional neural network, The first convolutional layer receives a single-channel input signal y(n) and generates a feature map with m channels. The middle convolutional layers maintain m channels, with a kernel size of i×i and padding of j. ReLU activation and normalization are applied after each convolutional layer to introduce nonlinear characteristics of the network. The last convolutional layer converts the feature map of m channels back to a single channel and outputs the initially recovered signal sequence y(t). 3b) The initially recovered signal sequence y(t) is input into the existing L bidirectional long short-term memory network modules, and residual skip connections are made between the first and last convolutional layers to obtain the enhanced signal sequence y(m); 3c) The enhanced signal sequence y(m) is input into the fully connected layer, and the recovered signal sequence is output after passing through the fully connected layer. That is, to complete the construction of the deep learning network model; The acquired time-domain nonlinear distortion dataset is input into the constructed deep learning network model for training to recover the time-domain distortion signal. ; For the recovered time-domain signal Channel estimation, equalization, and demodulation are performed to obtain the final bitstream b, which is implemented as follows: 5a) For the recovered signal The cyclic prefix (CP) is removed, followed by a Fast Fourier Transform (FFT) operation to obtain the frequency domain signal at the receiver. ; 5b) Extract the frequency domain signal pilot signal in The channel response matrix of the pilot signal is obtained by performing minimum linear mean square error (LMMSE) channel estimation. ; 5c) Channel response matrix to pilot signal Perform linear interpolation to obtain the channel response matrix of the entire signal. ; 5d) Channel response matrix The result X after equilibrium is obtained by performing minimum mean square error MMSE equilibrium. K Its implementation includes: 5d1) Using the average power of the transmitted signal, the average power of the noise, and the frequency domain channel response matrix Determine the MMSE equalizer coefficients : ， in, The average power of the noise, The average power of the transmitted signal, express The conjugate matrix of , where I is the identity matrix; 5d2) Using MMSE equalizer coefficients Calculate the equilibrium result X K : ， in, This refers to the frequency domain signal received at the receiver. 5e) Perform a sign decision on the equalized result X according to the constellation diagram rules, and demap the decision result to the corresponding bit data, that is, demodulate the bit data b.

2. The method according to claim 1, characterized in that, The acquisition of the dataset for time-domain distortion recovery of satellite communication OFDM signals is implemented as follows: 2a) In a satellite communication system, the transmitting end randomly generates a bit stream a, performs digital modulation on the data stream, performs inverse Fourier transform (IFFT) calculation on the modulated signal, adds a pilot signal and inserts a cyclic prefix (CP) into the time domain signal x(n) as a tag; 2b) The time-domain signal x(n) is amplified by the nonlinear power amplifier HPA, and the amplified signal z(n) is then transmitted through the satellite channel h with added noise. The output signal y(n) is then used as a sample; 2c) Use the obtained samples y(n) and labels x(n) to form a dataset.

3. The method according to claim 1, characterized in that, The process of inputting the acquired dataset into the constructed deep learning network model for training, validation, and testing to recover the time-domain distorted signal includes the following: 4a) Perform dimensionality transformation and normalization preprocessing on the temporal complex signals in the dataset to adapt to the network's input rules; 4b) Divide the preprocessed data into training, validation, and test sets in a ratio of 7:1.5:1.5; 4c) Using the mean squared error (MSE) as the loss function, calculate the error between the network output and the true label; 4d) Set the minimum batch size and learning rate. Training epochs and total number of samples N; 4e) Initialize the parameters in the neural network model to give the network an initial state; 4f) Input the partitioned training set into the constructed deep learning network model, start forward propagation, and output the initial results of the model; 4g) After each forward propagation, first calculate the MSE loss between the network output and the real label, and then calculate the gradient of the loss function with respect to the network parameters; 4h) The gradient is propagated from the output layer back to the input layer using the backpropagation algorithm. The Adam optimizer is used to update the network parameters step by step according to the calculated gradient until the set training epoch is reached to complete the training. 4i) Input the divided validation set and test set into the validated network model to verify the robustness of the model; 4j) Generate a new set of distorted data and input it into the network model after testing to obtain the recovered signal sequence. .

4. The method according to claim 1, characterized in that, In step 5b), the pilot signal from the receiver is extracted. By performing minimum linear mean square error (LMMSE) channel estimation, the channel response matrix of the pilot signal is obtained. Its implementation includes: 5b1) Assume the channel autocorrelation matrix It is a diagonal matrix of the real channel H; 5b2) Pilot signal for the receiving end Perform least squares estimation to obtain the initial estimate. : ， in This is the pilot signal for the transmitting end; 5b3) Set the constants for the modulation method ; 5b4) Based on the autocorrelation matrix and the results of the initial estimate Calculate the channel response matrix of the pilot signal. : ， Where SNR is the signal-to-noise ratio of the signal, and I is the identity matrix.

5. The method according to claim 1, characterized in that, The channel response matrix of the pilot signal in step 5c) Perform linear interpolation to obtain the channel response matrix of the entire signal. Its implementation includes: 5c1) Determine the channel response matrix of the pilot signal. The location of the missing data points; 5c2) Find the coordinates of the neighboring data points of the missing data point, and use the linear interpolation formula to calculate the y-value of the missing data point at x: ， Where (x, y) is the coordinate of a missing data point, and (x1, y1) and (x2, y2) are the coordinates of its neighboring points; 5c3) Fill in the coordinates of all the missing data points into the corresponding positions to obtain the channel response matrix of the entire signal. .

6. The method according to claim 1, characterized in that, In step 5e), the balanced result X K Symbol determination is performed according to constellation diagram rules to demodulate bit stream b. The implementation includes: 5e1) Equilibrium result X K Each signal in the diagram is mapped onto a constellation diagram corresponding to its modulation scheme, and the Euclidean distance between the signal and each point in the constellation diagram is calculated. ; ， in, Is it the equilibrium result X? K Each signal coordinate in the data, These are the coordinates of each point in the constellation chart; 5e2) Based on the distance calculation results, select the constellation point with the closest distance as the decision result; 5e3) Map the decision result point to the original bit stream b.

7. A device for restoring time-domain nonlinear distortion of OFDM signals for implementing the method of claim 1, characterized in that... ,include: The digital modulation module is used to digitally modulate the bit stream and map it into frequency domain complex symbols; The OFDM modulation module is used to convert frequency domain complex symbols into time domain signals through inverse fast Fourier transform (IFFT), while adding a cyclic prefix (CP) to reduce multipath interference. The power amplifier module is used to amplify the power of the OFDM modulated signal to ensure the transmission power of the signal in the satellite communication link; The upconversion and radio frequency module is used to convert the amplified baseband signal into a radio frequency signal and transmit it to the satellite through the antenna to realize wireless signal transmission; The receiving radio frequency and down-conversion module is used to receive radio frequency signals from satellites and down-convert them back to baseband signals in preparation for subsequent signal processing. The time-domain restoration network module is used to perform preliminary nonlinear distortion recovery processing on the received satellite signals; The OFDM demodulation module is used to perform Fast Fourier Transform (FFT) on the initially recovered signal to remove pilot symbols and cyclic prefixes. The channel estimation and equalization module is used to estimate channel characteristics and perform equalization processing to compensate for fading and distortion in the channel and restore the original transmitted signal. The digital demodulation module is used to demodulate the equalized signal into a bit stream, thus completing the data demodulation.

Images