An affine frequency division multiplexing peak-to-average power ratio reduction method and system based on deep learning

Abstract

The application discloses an affine frequency division multiplexing peak-to-average power ratio reduction method and system based on deep learning, and the method comprises the following steps: acquiring a simulation affine frequency division multiplexing transmitting signal; constructing a deep neural network model based on unsupervised learning, and training the deep neural network based on the simulation affine frequency division multiplexing transmitting signal to obtain a peak-to-average power ratio optimization model; acquiring the affine frequency division multiplexing transmitting signal and inputting the affine frequency division multiplexing transmitting signal into the peak-to-average power ratio optimization model for optimization to obtain an optimization value, and combining the optimization value with the affine frequency division multiplexing transmitting signal to reduce the peak-to-average power ratio. The application designs a method and system for reducing the peak-to-average power ratio of the affine frequency division multiplexing based on the deep learning technology, fills the blank of the deep learning in the reduction of the peak-to-average power ratio of the affine frequency division multiplexing, and enriches the application of the deep learning technology in the wireless communication system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to an affine frequency division multiplexing peak-to-average power ratio reduction method and system based on deep learning. BACKGROUND

[0002] With the development of technology, people's demand for wireless communication under high-speed movement is gradually increasing, such as high-speed rail, Internet of Vehicles, etc. Therefore, the next generation of wireless communication puts forward higher requirements for wireless communication under high-speed movement. In 4G and 5G wireless communication technology, orthogonal frequency division multiplexing has been widely applied due to its good anti-multipath fading ability, high spectrum utilization and excellent performance under linear time-invariant channel. However, in linear time-varying channel, the orthogonality of orthogonal frequency division multiplexing will decrease sharply, resulting in inter-carrier interference, which greatly reduces the performance of orthogonal frequency division multiplexing. In order to obtain good performance in linear time-varying channel and meet the demand of wireless communication under high-speed movement, many new signal modulation technologies have been proposed, and affine frequency division multiplexing (AFDM) is born in such a background.

[0003] AFDM is based on the generalization of discrete Fourier transform-discrete affine Fourier transform, which can realize full diversity because its linear frequency pulse parameters can adapt to channel characteristics, so as to realize complete delay Doppler representation of the channel in the discrete affine Fourier transform domain. Experiments show that AFDM has excellent performance under linear time-varying channel and can meet the demand of wireless communication under high-speed movement.

[0004] As the generalization of Orthogonal Frequency Division Multiplexing (OFDM), AFDM faces the same problem as OFDM, which is high Peak to Average Power Ratio (PAPR). High PAPR will make the signal easily enter the nonlinear region of the power amplifier, causing the signal to produce nonlinear distortion, resulting in the consequences of rising Bit Error Rate (BER), spectrum leakage, etc. In order to effectively reduce PAPR, the academic community has proposed many PAPR suppression algorithms, such as clipping, selective mapping (SLM), partial transmission sequence (PTS), etc. In recent years, with the vigorous development of computer computing power, deep learning technology has also developed rapidly. Deep learning technology is widely used in many fields, such as computer vision, natural language processing, etc. At the same time, deep learning technology also has a wide range of applications in the field of wireless communication, and its potential in the field of wireless communication is being continuously tapped. Inspired by this, the present application applies deep learning technology to the affine frequency division multiplexing system to reduce its PAPR, and simulation shows that the present application effectively reduces the PAPR with only a very small loss to the BER performance of the affine frequency division multiplexing communication system. SUMMARY

[0005] The present application aims to solve the deficiencies of the prior art and provides the following solutions:

[0006] An affine frequency division multiplexing peak-to-average power ratio reduction method based on deep learning, comprising the following steps:

[0007] Obtain a simulated affine frequency division multiplexing transmit signal;

[0008] Construct a deep neural network model based on unsupervised learning, and train the deep neural network based on the simulated affine frequency division multiplexing transmit signal to obtain a peak-to-average power ratio optimization model;

[0009] Obtain an affine frequency division multiplexing transmit signal and input it into the peak-to-average power ratio optimization model for optimization to obtain an optimized value, and combine the optimized value with the affine frequency division multiplexing transmit signal to reduce the peak-to-average power ratio.

[0010] Preferably, the simulated affine frequency division multiplexing transmit signal includes pilot and data;

[0011] The pilot is data known to both the transmitter and the receiver, and is used for channel estimation when receiving a signal;

[0012] The data carries the information to be transmitted.

[0013] Preferably, the deep neural network model includes: an input layer, several hidden layers, and an output layer;

[0014] The number of hidden layers is 8, the activation function of the hidden layers is ReLU, and the activation function of the output layer is Sigmoid;

[0015] Adam is used as the optimizer for the deep neural network model.

[0016] Preferably, the method for training the deep neural network includes:

[0017] To obtain the peak-to-average power ratio expression for RF multiplexing:

[0018]

[0019] Where, E{|s[n]| 2} represents the average power of the simulated radio frequency multiplexing transmit signal. s[n] represents the peak power of the simulated radio frequency multiplexing transmit signal, s[n] represents the time domain signal of the simulated radio frequency multiplexing transmit signal after discrete affine Fourier transform, N represents the number of linear frequency modulated subcarriers, x[m] represents the simulated radio frequency multiplexing transmit signal, and c1 and c2 represent two parameters of the simulated radio frequency multiplexing.

[0020] The peak-to-average power ratio (PAPR) expression is used as a loss function to train the deep neural network model, thereby obtaining the PAPR optimization model.

[0021] Preferably, the method for reducing the peak-to-average power ratio includes:

[0022] The simulated radio frequency multiplexing transmit signal is acquired and input into the peak-to-average power ratio optimization model for optimization to obtain optimized values, which include: guard interval insertion point and active constellation extension point;

[0023] The optimized value is combined with the simulated radio frequency multiplexing transmit signal to reduce the peak-to-average power ratio: the guard interval insertion point is embedded between the pilot and the data to prevent interference between the pilot and the data and reduce the peak-to-average power ratio; the active constellation extension point reduces the peak-to-average power ratio by finding the optimal value within the extension range.

[0024] The present invention also provides a deep learning-based simulated radio frequency multiplexing peak-to-average power ratio reduction system, wherein the system applies the method described in any of the above-mentioned methods and includes: a signal acquisition module, a model construction module, and a power ratio reduction module;

[0025] The signal acquisition module is used to acquire simulated radio frequency multiplexing transmission signals;

[0026] The model construction module constructs a deep neural network model based on unsupervised learning, and trains the deep neural network based on the simulation AFDM transmit signal, to obtain a PAPR optimization model;

[0027] The PAPR reduction module is configured to obtain an AFDM transmit signal and input the AFDM transmit signal into the PAPR optimization model for optimization, to obtain an optimized value, and combine the optimized value with the AFDM transmit signal to reduce PAPR.

[0028] Preferably, the simulation AFDM transmit signal includes a pilot and data.

[0029] The pilot is data known to both the transmitter and the receiver, and is used for channel estimation when receiving a signal.

[0030] The data carries information to be transmitted.

[0031] Preferably, the deep neural network model includes an input layer, a plurality of hidden layers, and an output layer.

[0032] The plurality of hidden layers is eight layers, the activation function of the hidden layers is ReLU, and the activation function of the output layer is Sigmoid.

[0033] Adam is used as the optimizer of the deep neural network model.

[0034] Preferably, the process of training the deep neural network includes:

[0035] The PAPR expression of AFDM is obtained:

[0036]

[0037] wherein E{|s[n]|2} represents the average power of the simulation AFDM transmit signal, 2 represents the peak power of the simulation AFDM transmit signal, s[n] represents a time-domain signal converted by discrete affine Fourier transform from the AFDM transmit signal, N represents the number of linear frequency modulation subcarriers, x[m] represents the simulation AFDM transmit signal, and c1 and c2 represent two parameters of AFDM.

[0038] The PAPR expression is used as a loss function to train the deep neural network model, to obtain the PAPR optimization model.

[0039] Preferably, the working process of the PAPR reduction module includes:

[0040] ​The affine frequency division multiplexing transmit signal is acquired and input into the peak-to-average power ratio optimization model for optimization to obtain an optimization value, which includes a guard interval insertion point and an active constellation extension point.

[0041] The optimization value is combined with the affine frequency division multiplexing transmit signal for peak-to-average power ratio reduction: the guard interval insertion point is embedded between the pilot and the data to prevent pilot-to-data interference and reduce the peak-to-average power ratio; and the active constellation extension point reduces the peak-to-average power ratio by finding an optimal value in an extension range.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] (1) The present application designs a method and system for reducing the peak-to-average power ratio of affine frequency division multiplexing based on deep learning technology, fills the gap of deep learning in reducing the peak-to-average power ratio of affine frequency division multiplexing, and enriches the application of deep learning technology in wireless communication systems.

[0044] (2) The present application innovatively uses a deep neural network to simultaneously optimize the guard interval insertion point and the active constellation extension point, greatly reduces the peak-to-average power ratio while keeping the system complexity almost unchanged.

[0045] (3) The present application greatly reduces the peak-to-average power ratio while having little effect on the bit error rate performance of the wireless communication system. At the same time, since the guard interval insertion point and the active constellation extension point used to optimize the peak-to-average power ratio do not carry any side information, the transmitting end does not need to transmit any additional data to the receiving end, so that the system does not cause additional spectrum loss. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings described in the following embodiments are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The method flowchart of the embodiment of the present application;

[0048] Figure 2 The symbol architecture diagram of the embodiment of the present application;

[0049] Figure 3 The optimization point constellation distribution diagram of the embodiment of the present application;

[0050] Figure 4This is a simulation comparison of the peak-to-average power ratio (PAPR) performance of the simulated radio frequency multiplexed signal after optimization of PAPR and the original signal according to an embodiment of the present invention.

[0051] Figure 5 This is a simulation comparison of the bit error rate performance of the simulated radio frequency multiplexing signal and the original simulated radio frequency multiplexing signal under existing channel estimation and symbol detection methods according to an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1

[0055] In this embodiment, as Figure 1 As shown, a deep learning-based method for reducing the peak-to-average power ratio of radio frequency demultiplexing includes the following steps:

[0056] S1. Obtain the simulated radio frequency multiplexing transmit signal.

[0057] The simulated radio frequency multiplexing (RFD) transmit signal includes: pilots and data; the pilots are data known to both the transmitter and receiver and are used to estimate the channel when receiving signals; the data carries the information to be transmitted.

[0058] In this embodiment, the simulated radio frequency multiplexing transmit signal architecture is as follows: Figure 2 As shown. In the original transmission architecture, the pilot signal is known data to both the transmitter and receiver, used for channel estimation when receiving signals; the data carries the information to be transmitted; the guard interval is an empty string of length Q located at both ends of the pilot signal, used to prevent interference between the pilot signal and the data, where Q = (2α) max +1)(l max +1)-1, α max With l max These are the normalized maximum Doppler frequency shift and time delay, respectively.

[0059] S2. Construct a deep neural network model based on unsupervised learning, and train the deep neural network based on simulated radio frequency multiplexing transmission signals to obtain a peak-to-average power ratio optimization model.

[0060] The deep neural network model includes an input layer, several hidden layers, and an output layer; there are 8 hidden layers, with ReLU as the activation function for the hidden layers and Sigmoid as the activation function for the output layer; Adam is used as the optimizer for the deep neural network model.

[0061] Methods for training deep neural networks include: obtaining the peak-to-average power ratio expression for simulated radio frequency multiplexing:

[0062]

[0063] Where, E{|s[n]| 2} represents the average power of the simulated radio frequency multiplexing transmit signal. The peak power of the simulated radio frequency multiplexing (RFD) transmit signal is represented by s[n], the time-domain signal of the simulated RFD transmit signal after discrete affine Fourier transform is represented by s[n], N represents the number of linear frequency modulated subcarriers is represented by N, x[m] represents the simulated RFD transmit signal is represented by x[m], and c1 and c2 represent two parameters of the simulated RFD; in this embodiment, c1 = (2α) max +1) / 2N, where c2 is any irrational number or a rational number much smaller than 1 / 2N. The peak-to-average power ratio (PAPR) expression is used as the loss function to train a deep neural network model, resulting in an optimized PAPR model.

[0064] S3. Obtain the simulated radio frequency multiplexing transmission signal and input it into the peak-to-average power ratio optimization model for optimization, obtain the optimized value, and combine the optimized value with the simulated radio frequency multiplexing transmission signal to reduce the peak-to-average power ratio.

[0065] Methods for reducing peak-to-average power ratio (PAPR) include: acquiring a simulated radio frequency multiplexing (RFD) transmit signal and inputting it into a PAPR optimization model for optimization to obtain optimized values, including guard interval insertion points and active constellation extension points; and combining the optimized values ​​with the simulated RFD transmit signal to reduce PAPR: the guard interval insertion point is embedded between the pilot and data to prevent interference between pilot and data and reduce PAPR; the active constellation extension point reduces PAPR by finding the optimal value within the extension range.

[0066] In this embodiment, pilot signals and data are used as inputs to a neural network to optimize the peak-to-average power ratio (PAPR), outputting optimized values ​​for reducing PAPR, including guard interval insertion points and active constellation extension points. In the optimized transmission signal architecture, the guard interval insertion point is located at the guard interval position in the original architecture, preventing interference between pilot signals and data while optimizing PAPR; the data after active constellation extension replaces the original data, optimizing PAPR without changing the transmitted bit information or reducing the minimum Euclidean distance. In this embodiment, Quadrature Phase Shift Keying (QPSK) is used as an example, and the constellation distribution of the optimization points for reducing PAPR is as follows: Figure 3 As shown. For the guard interval optimization point, its constellation range is a circle with radius R. Increasing R can improve the peak-to-average power ratio reduction performance, but due to the increased energy of the guard interval insertion point, its interference with pilots and data will increase, thereby reducing the bit error rate performance during symbol detection. Therefore, choosing an appropriate R to balance the bit error rate and peak-to-average power ratio reduction performance is crucial. For the active constellation extension point, its constellation range extends outward from the original data constellation point. To prevent the energy of the active constellation extension point from being too large, we set the upper limit of the normalized amplitude extension of both its real and imaginary parts to 1. Since this extension range ensures that the minimum Euclidean distance between constellation points does not decrease, it can ensure that the bit error rate performance is not affected to the greatest extent.

[0067] This embodiment optimizes the peak-to-average power ratio (PAPR) of simulated radio frequency multiplexing (RFD) transmission signals by deploying a pre-trained deep neural network at the signal transmitting end. This embodiment comprises two stages: an offline training stage and an online deployment stage. In the offline training stage, a corresponding deep neural network is constructed based on the architecture of the pilot signals, data, and PAPR optimization points. A loss function for reducing PAPR is applied to this network, and the optimal parameter combination is found by continuously adjusting various neural network parameters. Next, a large number of simulated RFD transmission signals (i.e., pilot signals and data) are used to form the training dataset for the neural network, and input into the neural network for training until optimal performance is achieved. In the online deployment stage, the trained deep neural network model is deployed at the transmitting end. The user inputs the simulated RFD signal to be optimized into the pre-trained neural network, which outputs optimization points for reducing PAPR. Combining the transmitted pilot signals and data with these optimization points generates a simulated RFD transmission signal with low PAPR.

[0068] Example 2

[0069] In the embodiment, an affine frequency division multiplexing peak-to-average power ratio reduction system based on deep learning comprises a signal acquisition module, a model construction module and a power ratio reduction module.

[0070] The signal acquisition module is configured to acquire a simulation affine frequency division multiplexing transmission signal.

[0071] The simulation affine frequency division multiplexing transmission signal comprises a pilot and data.

[0072] The model construction module is configured to construct a deep neural network model based on unsupervised learning, and train the deep neural network based on the simulation affine frequency division multiplexing transmission signal to obtain a peak-to-average power ratio optimization model.

[0073] The deep neural network model comprises an input layer, a plurality of hidden layers and an output layer.

[0074] The process of training the deep neural network comprises obtaining a peak-to-average power ratio expression of the affine frequency division multiplexing:

[0075]

[0076] wherein E{|s[n]|2} represents the average power of the simulation affine frequency division multiplexing transmission signal, 2 and E{|s[n]|2} represents the peak power of the simulation affine frequency division multiplexing transmission signal. s[n] represents a time domain signal converted by the simulation affine frequency division multiplexing transmission signal through a discrete affine Fourier transform, N represents the number of linear frequency modulation subcarriers, x[m] represents the simulation affine frequency division multiplexing transmission signal, and c1 and c2 represent two parameters of the affine frequency division multiplexing. The peak-to-average power ratio expression is used as a loss function to train the deep neural network model to obtain the peak-to-average power ratio optimization model.

[0077] The power ratio reduction module is configured to acquire the simulation affine frequency division multiplexing transmission signal and input the simulation affine frequency division multiplexing transmission signal into the peak-to-average power ratio optimization model for optimization to obtain an optimized value.

[0078] The working process of the power ratio reduction module includes: obtaining an affine frequency division multiplexing transmission signal and inputting the signal into a peak-to-average power ratio optimization model for optimization to obtain an optimization value, the optimization value including: a guard interval insertion point and an active constellation expansion point; and combining the optimization value with the affine frequency division multiplexing transmission signal to reduce the peak-to-average power ratio: the guard interval insertion point is embedded between the pilot and the data, for preventing interference between the pilot and the data and reducing the peak-to-average power ratio; and the active constellation expansion point reduces the peak-to-average power ratio by finding an optimal value in an expansion range.

[0079] Embodiment three

[0080] In order to illustrate the technical progressiveness of the present application, the peak-to-average power ratio and the bit error rate performance of the affine frequency division multiplexing peak-to-average power ratio reduction method based on deep learning proposed in the present embodiment are simulated on a Matlab platform.

[0081] The parameter settings of the simulation are as follows: the carrier frequency and bandwidth are 4 GHz and 32 kHz respectively; the number of linear frequency modulation subcarriers N = 32; the maximum time delay of the channel is 62.5 μs, and the normalized maximum time delay is 2; the maximum Doppler shift of the channel is 1 kHz, and the normalized maximum Doppler shift is 1; the length of the guard interval is 2Q = 2((2α max +1)(l max +1)-1) = 16; the number of channels P = 2; the complex gain of the channel is generated by independent complex random variables with zero mean and 1 / P variance; the normalized time delay of the channel is l = [0, 2]; the normalized Doppler shift of the channel at the i-th channel is α i = α max cosθ i , where α max represents the maximum Doppler shift, and θ i is uniformly distributed on [-π, π]; the ratio of the pilot signal-to-noise ratio to the data signal-to-noise ratio SNRp / SNRd = 20 dB. In terms of the setting of the neural network parameters, the neural network includes an input layer, a hidden layer and an output layer, and the neural network has a total of 8 hidden layers; the activation function of the neural network hidden layer is ReLU, the activation function of the output layer is Sigmoid, and Adam is used as the optimizer of the neural network. In terms of the peak-to-average power ratio optimization parameters, the distribution radius of the guard interval insertion point is R = 0.2. Specifically, Figure 4 The peak-to-average power ratio performance of the method is shown by comparing the complementary cumulative distribution function (CCDF), where CCDF(PAPR0) = prob(PAPR > PAPR0), i.e. the probability that the peak-to-average power ratio of the transmitted signal sample is greater than PAPR0; Figure 5 The bit error rate performance of the method is shown.Figure 4 Compared with the CCDF curve of the original AF-OFDM signal; Figure 5 The existing embedded channel estimation method and the Minimum Mean Square Error (MMSE) symbol detection method are adopted, and the bit error rate performance is compared with that of the original AF-OFDM signal.

[0082] Through simulation, it can be seen that, compared with the original AF-OFDM signal, the signal optimized by the deep neural network can effectively reduce the peak-to-average power ratio. In the simulation process, we tried to optimize only the guard interval insertion point and simultaneously optimize the guard interval insertion point and active constellation expansion point, as shown in Figure 4 The scheme of simultaneously optimizing the guard interval insertion point and the active constellation expansion point can achieve a lower peak-to-average power ratio than the scheme of optimizing only the guard interval insertion point. It is worth noting that, since there is no significant difference in the network architecture of the two optimization methods, the complexity of the two optimization methods is almost equal, which means that simultaneously optimizing the guard interval insertion point and the active constellation expansion point can bring higher peak-to-average power ratio reduction efficiency. At the same time, the signal optimized by the deep neural network still performs very well in bit error rate. As shown in Figure 5 The scheme of optimizing only the guard interval insertion point can achieve almost the same performance as the original signal; the bit error rate performance of the scheme of simultaneously optimizing the guard interval insertion point and the active constellation expansion point has decreased compared with the scheme of optimizing only the guard interval insertion point, but still maintains good performance. In summary, if a system has a higher demand for peak-to-average power ratio reduction, the scheme of simultaneously optimizing the guard interval insertion point and the active constellation expansion point can be adopted; if a system needs to reduce the peak-to-average power ratio to a certain extent, but has a more stringent requirement for the bit error rate performance of the system, the scheme of optimizing only the guard interval insertion point is more suitable.

[0083] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. An affine frequency division multiplex peak-to-average power ratio reduction method based on deep learning, characterized in that, The method comprises the following steps: obtain a simulation OFDM signal; construct a deep neural network model based on unsupervised learning, and train the deep neural network based on the simulation OFDM signal to obtain a PAPR optimization model; obtain an OFDM signal and input it into the PAPR optimization model for optimization to obtain an optimization value, and combine the optimization value with the OFDM signal to reduce PAPR; the simulation OFDM signal comprises pilot and data; the pilot is known data between the transmitter and the receiver, and is used for channel estimation when receiving a signal; the data carries information to be transmitted; the deep neural network model comprises an input layer, a plurality of hidden layers, and an output layer; the plurality of hidden layers are 8 layers, the activation function of the hidden layer is ReLU, and the activation function of the output layer is Sigmoid; Adam is used as the optimizer of the deep neural network model; the method for training the deep neural network comprises: obtain a PAPR expression of OFDM: wherein E{|s[n]|2} represents the average power of the simulated AFDM transmitting signal, 2} represents the average power of the simulated AFDM transmitting signal, represents the peak power of the simulated AFDM transmitting signal, s[n] represents the time domain signal converted from the AFDM transmitting signal by the discrete affine Fourier transform, N represents the number of the linear frequency modulation subcarriers, x[m] represents the simulated AFDM transmitting signal, and c1 and c2 represent two parameters of the AFDM. use the PAPR expression as a loss function to train the deep neural network model to obtain the PAPR optimization model; the method for reducing PAPR comprises: obtain an OFDM signal and input it into the PAPR optimization model for optimization to obtain an optimization value, the optimization value comprising a guard interval insertion point and an active constellation expansion point; and combine the optimization value with the OFDM signal to reduce PAPR: the guard interval insertion point is embedded between the pilot and the data to prevent pilot data interference and reduce PAPR; the active constellation expansion point reduces PAPR by finding the optimal value in the expansion range.

2. An affine frequency division multiplex peak-to-average power ratio reduction system based on deep learning, the system applying the method of claim 1, characterized in that, comprise: a signal acquisition module, a model construction module, and a power ratio reduction module; the signal acquisition module is used to obtain a simulation OFDM signal; the model construction module constructs a deep neural network model based on unsupervised learning, and trains the deep neural network based on the simulation OFDM signal to obtain a PAPR optimization model; the power ratio reduction module is used to obtain an OFDM signal and input it into the PAPR optimization model for optimization to obtain an optimization value, and combine the optimization value with the OFDM signal to reduce PAPR.

3. The deep learning-based simulated radio frequency multiplexing peak-to-average power ratio reduction system according to claim 2, characterized in that, the simulation OFDM signal comprises pilot and data; the pilot is known data between the transmitter and the receiver, and is used for channel estimation when receiving a signal; the data carries information to be transmitted.

4. The system according to claim 2, wherein the system is a deep learning based system for reducing peak-to-average power ratio (PAPR) of an OFDM signal. the deep neural network model comprises an input layer, a plurality of hidden layers, and an output layer; the plurality of hidden layers are 8 layers, the activation function of the hidden layer is ReLU, and the activation function of the output layer is Sigmoid; Adam is used as the optimizer of the deep neural network model.

5. The system according to claim 2, wherein the system is a deep learning based system for reducing peak-to-average power ratio (PAPR) of an OFDM signal. the process for training the deep neural network comprises: obtain a PAPR expression of OFDM: wherein E{s[n]} represents the average power of the simulated AFDM transmit signal, 2} represents the average power of the simulated AFDM transmit signal, represents the peak power of the simulated AFDM transmit signal, s[n] represents the time domain signal converted from the AFDM transmit signal by the discrete affine Fourier transform, N represents the number of the linear frequency modulation subcarriers, x[m] represents the simulated AFDM transmit signal, and c1 and c2 represent two parameters of the AFDM. The peak-to-average power ratio expression is taken as a loss function to train the deep neural network model to obtain the peak-to-average power ratio optimization model.

6. The system according to claim 3, wherein the system is a deep learning based system for reducing peak-to-average power ratio (PAPR) of an OFDM signal. The working process of the power ratio reduction module includes: An affine frequency division multiplexing transmission signal is obtained and input into the peak-to-average power ratio optimization model for optimization to obtain an optimization value, which includes a guard interval insertion point and an active constellation extension point. The optimization value is combined with the affine frequency division multiplexing transmission signal to reduce the peak-to-average power ratio. The guard interval insertion point is embedded between the pilot and the data to prevent pilot-to-data interference and reduce the peak-to-average power ratio; and the active constellation extension point reduces the peak-to-average power ratio by finding an optimal value in an extension range.

Images