CP-free PS-OFDM communication system based on autoencoder

Abstract

The present invention provides a CP-free PS-OFDM communication system based on an autoencoder, comprising a transmitter, a channel and a receiver; the transmitter comprises a modulation module, a serial-to-parallel conversion module, an inverse Fourier transform module and a transmitter pulse shaping module connected in sequence, the transmitter pulse shaping module comprising a transmitting DNN, the transmitting DNN comprising a transmitting input layer connected to the output end of the inverse Fourier transform module, and a transmitting output layer; the channel is a multipath fading channel; the receiver comprises a receiving pulse shaping module, a coarse equalization module and a demodulation mapping module connected in sequence, the receiving pulse shaping module comprising a receiving DNN, the receiving DNN comprising a receiving input layer for inputting a received signal, and a receiving output layer, the receiving output layer being a receiving PS filter. The present invention can reduce the influence of inter-symbol interference and inter-carrier interference on signal detection, control the time and frequency characteristics of the OFDM signal, and realize flexible configuration of pulse shaping to flexibly support different communication scenarios.

Description

Technical Field

[0001] The present invention relates to a CP-free PS-OFDM communication system based on an automatic encoder. Background Art

[0002] In the future development of mobile communication technology, achieving higher spectral efficiency and a wider range of application scenarios places higher demands on modulation techniques. Orthogonal Frequency Division Multiplexing (OFDM) effectively addresses selective fading and narrowband interference by converting frequency-selective fading channels into multiple narrowband flat-fading channels, achieving efficient broadband transmission. Furthermore, traditional OFDM systems reduce inter-symbol interference (ISI) and inter-carrier interference (ICI) by adding a cyclic prefix (CP) of a certain length. However, the introduction of CP indirectly reduces spectral efficiency. Traditional CP-based OFDM systems (CP-OFDM systems) use a fixed rectangular window for windowing, which not only has significant drawbacks in suppressing out-of-band power leakage and spectral efficiency (SE) but also results in a single pulse shape, making it inflexible in supporting multiple communication scenarios.

[0003] Therefore, there is an urgent need to find a communication system that can reduce ISI and ICI, control the time and frequency characteristics of OFDM signals, and avoid the single pulse shape defect. Summary of the Invention

[0004] The purpose of the present invention is to propose a CP-free PS-OFDM communication system based on an autoencoder, which can reduce the impact of inter-symbol interference and inter-carrier interference on signal detection, control the time and frequency characteristics of OFDM signals, and realize flexible configuration of pulse shaping to flexibly support different communication scenarios.

[0005] The present invention is achieved through the following technical solutions:

[0006] A CP-free PS-OFDM communication system based on an autoencoder, including a transmitter, a channel, and a receiver;

[0007] The transmitter includes a modulation module, a serial-to-parallel conversion module, an inverse Fourier transform module and a transmitter pulse shaping module connected in sequence. The transmitter pulse shaping module includes a transmitting DNN. The transmitting DNN includes a transmitting input layer connected to the output end of the inverse Fourier transform module, and a transmitting output layer. The transmitting output layer is a transmitting PS filter. The baseband modulated signal output by the modulation module is converted into a parallel frequency domain signal by the serial-to-parallel conversion module. The parallel frequency domain signal is converted into a time domain signal by the inverse Fourier transform module. The time domain signal is converted into a transmitting signal by the transmitting pulse shaping module. Where N is the number of subcarriers, X(k,i) is the frequency domain symbol corresponding to the i-th OFDM signal of the k-th subcarrier, and the PS filter at the transmitter is g k,i (n), g tx (n) is the pulse shape at the transmitting end, T is the duration of the discrete time domain OFDM signal, and n represents the nth sampling point in the discrete time domain;

[0008] The channel is a multipath fading channel;

[0009] The receiver includes a receiving end pulse shaping module, a coarse equalization module and a demodulation mapping module connected in sequence. The receiving end pulse shaping module includes a receiving DNN. The receiving DNN includes a receiving input layer for inputting the received signal and a receiving output layer. The receiving output layer is a receiving end PS filter. The received signal passes through the receiving end pulse shaping module to obtain a time domain received signal. In the coarse equalization module, the time domain received signal and the time domain channel are FFT-ed, and the two frequency domain information obtained are divided correspondingly. The division result is input into the demodulation mapping module. The demodulation mapping module is a neural network composed of multiple sub-neural networks. Each sub-neural network is implemented by DNN. The output of each sub-neural network is cascaded and spliced ​​to form the output of the demodulation mapping module.

[0010] Furthermore, the transmitting DNN also includes a transmitting hidden layer consisting of a layer of N neurons, a layer of 2N neurons and a power normalization layer, and the receiving DNN also includes a receiving hidden layer consisting of a layer of 2N neurons, a layer of N neurons and a power normalization layer.

[0011] Furthermore, the time domain received signal is Among them, y u To receive the signal, r k,i (n) is the PS filter at the receiving end, γ rx (n) is the pulse shape at the receiving end.

[0012] Furthermore, in order to fit the ISI term in the received signal, at least two transmission signals are continuously transmitted into the channel, and the parameters and structures of the transmission DNN used by the at least two transmission signals are exactly the same.

[0013] Furthermore, the modulation module is constellation modulation to map the bit stream into frequency domain constellation symbols.

[0014] Furthermore, the multipath fading channel is Where τ represents the maximum delay path number, δ(·) represents the discrete impulse response, and a p is the complex baseband channel coefficient of the pth delay path, τ p represents the delay of the pth delay path.

[0015] Furthermore, the demodulation mapping module is composed of N / K sub-neural networks, and each sub-neural network is responsible for demodulating the signal on a subcarrier.

[0016] Furthermore, the demodulation mapping module includes an input layer connected to the coarse equalization module, an output layer formed by splicing the Softmax outputs of each sub-neural network, and a hidden layer, wherein the hidden layer consists of a layer of 2K neurons, a layer of 4K neurons, a layer of 8K neurons, and a layer of 2K neurons.

[0017] Furthermore, the loss function of the demodulation mapping module is defined as in, corresponds to the output of the k-th Softmax classifier, indicating the probability of symbol demodulation on the k-th subcarrier, and M represents the modulation order of the modulation module, P k,1 ,...,P k,M Indicates the one-hot encoding used by the label, is the expected function.

[0018] Furthermore, when training the transmitter, channel, and receiver, forward propagation training is used to calculate the loss function, and the gradient descent algorithm is used to complete back propagation and update the parameters until the training is completed.

[0019] The present invention has the following beneficial effects:

[0020] 1. The transmitting DNN of the transmitter pulse shaping module, the channel, and the receiving DNN of the receiver pulse shaping module constitute an autoencoder, which introduces a sophisticated pulse shaping mechanism within the framework of the traditional OFDM system, thereby optimizing the transmission waveform to adapt to the complex wireless communication environment, adjusting the energy distribution of the pulse in the time domain and frequency domain, and being able to better control and balance inter-symbol interference and inter-carrier interference. The powerful fitting ability of the DNN is used to optimize the detection of CP-free OFDM signals, effectively solving the problem of decreased spectrum efficiency caused by CP in the existing CP-OFDM system, and ultimately achieving an effective improvement in the spectrum efficiency of the communication system without affecting the BER performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Figure 1 Schematic diagram of the system model of the present invention.

[0023] Figure 2 Schematic diagram of the model structure of the DNN network of the present invention.

[0024] Figure 3 It is a structural diagram of the demodulation mapping module of the present invention.

[0025] Figure 4 Schematic diagram of the structure of the sub-neural network in the regulation mapping module of the present invention.

[0026] Figure 5 The figure is a comparison of the bit error rate performance between the present invention and the CP-OFDM system when the number of subcarriers is 64, the maximum number of delay paths is 16, and QPSK modulation is adopted.

[0027] Figure 6 The figure is a comparison of the bit error rate performance between the present invention and the CP-OFDM system when the number of subcarriers is 64, the maximum number of delay paths is 16, and QPSK modulation is adopted.

[0028] Figure 7 The figure is a comparison of the bit error rate performance between the present invention and the CP-OFDM system when the number of subcarriers is 256, the maximum number of delay paths is 18, and 16QAM modulation is adopted.

[0029] Figure 8 The figure is a comparison of the bit error rate performance between the present invention and the CP-OFDM system when the number of subcarriers is 256, the maximum number of delay paths is 18, and 16QAM modulation is adopted.

[0030] Among them, 1. Transmitter; 11. Modulation module; 12. Serial-to-parallel conversion module; 13. Inverse Fourier transform module; 14. Transmitter pulse shaping module; 2. Channel; 3. Receiver; 31. Receiver pulse shaping module; 32. Coarse equalization module; 33. Demodulation mapping module. DETAILED DESCRIPTION

[0031] like Figure 1 As shown, the CP-free PS-OFDM communication system based on the autoencoder includes a transmitter 1, a channel 2 and a receiver 3.

[0032] Transmitter 1 is used to convert the bit stream into an actual time-domain transmission signal and includes a modulation module 11, a serial-to-parallel conversion (S / P) module 12, an inverse Fourier transform (IFFT) module 13, and a transmitter pulse shaping (PS) module 14, which are connected in sequence. Modulation module 11 uses constellation modulation to map the bit stream into frequency-domain constellation symbols. Different modulation modes can be selected. In this embodiment, the QPSK modulation mode is selected. Modulation module 11 converts continuous 2-bit information into a single frequency-domain symbol. The frequency-domain symbol value range is {1+1j, 1-1j, -1+1j, -1-1j}. The S / P module 12 is used to convert the serial frequency-domain signal into a parallel frequency-domain signal. The IFFT module 13 is used to convert the baseband modulated signal into a time-domain signal and output it to the transmitter PS module to generate the final transmission signal.

[0033] The transmitter pulse shaping module 14 is used to perform pulse shaping on the time-frequency signal, including transmitting DNN (Deep Neural Network, deep neural network), such as Figure 2 As shown, the transmitting DNN includes a transmitting input layer, a transmitting output layer, and a transmitting hidden layer connected to the output end of the inverse Fourier transform module 13. The transmitting input layer is an OFDM time-frequency signal, the transmitting output layer is a transmitting end PS filter, and the transmitting hidden layer is composed of a layer of N neurons, a layer of 2N neurons and a power normalization layer, wherein there are no trainable parameters in the power normalization layer.

[0034] The baseband modulated signal output by the modulation module 11 is converted into a parallel frequency domain signal by the serial-to-parallel conversion module 12. The parallel frequency domain signal is converted into a time domain signal by the inverse Fourier transform module 13. The time domain signal is converted into a transmission signal by the transmitter pulse shaping module 14. Where N is the number of subcarriers, X(k,i) is the independent and identically distributed frequency domain symbol corresponding to the i-th OFDM signal of the k-th subcarrier (i.e., the output signal of the modulation module 11), and the transmitting end PS filter is g k,i (n), g k,i (n) can be regarded as the pulse shape g at the transmitter tx The time-frequency shifted version of (n), that is, g tx (n) is the pulse shape at the transmitting end, T is the duration of the discrete time domain OFDM signal, and since CP is not added in the present invention, T=N, and n represents the nth sampling point in the discrete time domain.

[0035] Channel 2 is a multipath fading channel. The multipath fading channel can be regarded as a channel layer without any optimized parameters. There are no internal training parameters, and it only realizes the conversion between the transmitted signal and the received signal. Specifically, the multipath fading channel is modeled as Where τ represents the maximum delay path number, δ(·) represents the discrete impulse response, and ap is the complex baseband channel coefficient of the pth delay path, τ p represents the delay of the pth delay path.

[0036] For this channel, the conversion between the transmitted signal x and the received signal y can be expressed as:

[0037] y u =H1x u +H2x u-1 +w u

[0038]

[0039] , where x u and x u-1 Represent the time domain transmitted signals at the current moment and the previous moment respectively, that is, for the received signal y at the current moment u In the case of channel 2, the tail of the previous signal due to the multipath fading channel will produce ISI. In order to fit the ISI term in the received signal, two transmit signals are continuously transmitted into channel 2. The parameters and structure of the transmit DNN used in these two transmit signals are exactly the same.

[0040] The receiver 3 includes a receiving-end pulse shaping module 31, a coarse equalization module 32, and a demodulation mapping module 33 connected in sequence. The receiving-end pulse shaping module 31 is used to perform pulse shaping on the received time-domain received signal, and together with the transmitting-end pulse shaping, eliminates ISI and ICI. The receiving DNN includes a receiving input layer for inputting the received signal, a receiving output layer, and a receiving hidden layer. The receiving output layer is a receiving-end PS filter. The receiving hidden layer consists of a layer of 2N neurons, a layer of N neurons, and a power normalization layer.

[0041] The received signal is converted into a time domain received signal after passing through the receiving end pulse shaping module 31. Among them, y u To receive the signal, r k,i (n) is the PS filter at the receiving end, , γ rx (n) is the pulse shape at the receiving end,

[0042] In the coarse equalization module 32, the time domain received signal and the time domain channel 2 are subjected to FFT, and the two frequency domain information obtained are correspondingly divided, thereby achieving coarse equalization, and the division result is input to the demodulation mapping module 33. More specifically, for the current moment received signal y u Perform N-point FFT on both channel 1 and channel 2 in the time domain, and perform corresponding point division on the two obtained N-point sequences to obtain the frequency domain information after coarse equalization.

[0043] like Figure 4 As shown, the demodulation mapping module 33 is implemented using a large neural network, specifically a neural network composed of N / K sub-neural networks, each of which is responsible for demodulating the signal on the subcarrier. Each sub-neural network is implemented by a DNN, and the output of each sub-neural network is cascaded and spliced ​​to form the output of the demodulation mapping module 33. The structural diagram of the sub-neural network is shown in FIG. Figure 4 shown.

[0044] More specifically, the demodulation mapping module 33 includes an input layer connected to the coarse equalization module 32, an output layer formed by concatenating the Softmax outputs of each sub-neural network, and a hidden layer. The hidden layer consists of a layer of 2K neurons, a layer of 4K neurons, a layer of 8K neurons, and a layer of 2K neurons. The first layer in the hidden layer and the last layer in the hidden layer use jump connections to prevent the gradient from disappearing.

[0045] The structures of transmitter 1, channel 2, and receiver 3 are built using Tensorflow 2.1 and Kreas 2.3.1. The training of transmitter 1, channel 2, and receiver 3 includes the following steps:

[0046] Step 1: Initialize network parameters, input labels and data sets, set relevant hyperparameters, and select appropriate loss functions. The data set contains 10 6 80% of the dataset is used as training set and 20% as validation set; the label is the one-hot encoding obtained by converting the real bit stream information, that is, every 2 m bits are converted to the corresponding 2 M The unique hot encoding of M represents the modulation order of the modulation module 11; the cross entropy is selected as the loss function, that is, the loss function of the demodulation mapping module 33 is defined as in, corresponds to the output of the k-th Softmax classifier, indicating the probability of symbol demodulation on the k-th subcarrier, and M represents the modulation order of the modulation module 11, P k,1 ,...,P k,M Indicates the one-hot encoding used by the label, is the expected function;

[0047] Step 2: Forward propagation training, calculate the loss function, and select the appropriate gradient descent algorithm to complete the back propagation and update the parameters until the training is completed;

[0048] Specifically, the Adam optimization algorithm and small batch data processing method are used to complete the training of the autoencoder, with batch size = 1500 and initial learning rate lr = 0.001;

[0049] Step 3: Calculate the loss curve and the accuracy fitting curve, and pay attention to the training time. If it does not converge, return to step 1 to adjust the hyperparameters and retrain. If it converges, proceed to step 4.

[0050] Specifically, a learning rate adjustment strategy is adopted in which the learning rate decays every 5 epochs, and the learning rate decay rate dl = 0.8;

[0051] Step 4: Input the test set sequence data to complete the model test. The test set contains 10 5 The model performance is tested by statistically analyzing the BER in the test set.

[0052] In both the present invention and the comparative solution (CP-OFDM system), the receiving end has perfect knowledge of Channel 2 information to ensure fair comparison. When using QPSK modulation or other signal modulation schemes, this OFDM signal detection model can be equivalently transferred to achieve OFDM signal detection under different modulation schemes.

[0053] In order to demonstrate the ISI / ICI interference cancellation performance of the present invention without CP, two groups of performance simulations were performed.

[0054] Scenario 1: Figure 5 and Figure 6 The simulation conditions are as follows: the number of system subcarriers is 64, the maximum delay path number MaxL is 16, QPSK modulation is adopted, that is, M=2, the number of subcarriers demodulated by the sub-neural network K=8, and training is carried out in the signal-to-noise ratio range of 18-25dB.

[0055] Scenario 2: Figure 7 and Figure 8 The simulation conditions are that the number of system subcarriers is 256, the maximum delay path number MaxL is 18, 16QAM modulation is used, that is, M=4, the number of subcarriers demodulated by the sub-neural network K=16, and training is carried out in the signal-to-noise ratio range of 23-30dB.

[0056] The comparison scheme uses the classic rectangular window CP-OFDM, which considers the case where the CP length exceeds the maximum delay and the case where there is no CP, and the received signal adopts single-tap equalization based on the perfect channel 2 information. However, the present invention does not use CP.

[0057] Figure 5 and Figure 7 The bit error rate performance of the PS-OFDM communication system based on the autoencoder is demonstrated. Figure 5 and Figure 7As shown by the red and blue lines in the middle, the CP length significantly impacts the performance of traditional OFDM equalization and demodulation. When the CP length exceeds the maximum delay, CP-OFDM demodulation is unaffected by ISI and ICI, resulting in excellent demodulation BER performance. Without CP, ISI and ICI caused by multipath effects cannot be eliminated, and the demodulation BER no longer decreases when SNR exceeds 20dB. Both figures show that, without requiring a CP, the present invention (PS-OFDM No CP) achieves an order of magnitude improvement in BER performance compared to the traditional CP-OFDM system without CP (CP-OFDM No CP), while maintaining minimal performance degradation compared to the traditional CP-OFDM system with a full CP (CP-OFDM Full CP).

[0058] Figure 6 and Figure 8 The spectral efficiency of the present invention is compared with that of a traditional CP-OFDM system. According to the 3GPP protocol, CP-OFDM is divided into two scenarios: Extended CP (Scenario 1) and Normal CP (Scenario 2). Because the present invention (PS-OFDM No CP) can achieve similar bit error rate performance without CP, compared to a traditional CP-OFDM system with full CP (CP-OFDM Full CP), the communication system proposed in this invention can achieve spectral efficiency gains of 25% and 7%, respectively.

[0059] The calculation formula for spectrum efficiency is: Among them, N cp is the length of the used CP, R is the code rate, and in this embodiment, R is 1, that is, channel coding is not considered.

[0060] The above description is merely a preferred embodiment of the present invention and therefore cannot be used to limit the scope of the present invention. In other words, equivalent changes and modifications made according to the scope of the patent application and the contents of the specification should still fall within the scope of the patent of the present invention.

Claims

1. A CP-free PS-OFDM communication system based on an autoencoder, characterized by: Includes transmitter, channel and receiver; The transmitter includes a modulation module, a serial-to-parallel conversion module, an inverse Fourier transform module and a transmitter pulse shaping module connected in sequence. The transmitter pulse shaping module includes a transmitting DNN. The transmitting DNN includes a transmitting input layer connected to the output end of the inverse Fourier transform module, and a transmitting output layer. The transmitting output layer is a transmitting PS filter. The baseband modulated signal output by the modulation module is converted into a parallel frequency domain signal by the serial-to-parallel conversion module. The parallel frequency domain signal is converted into a time domain signal by the inverse Fourier transform module. The time domain signal is converted into a transmitting signal by the transmitting pulse shaping module. Where N is the number of subcarriers, X(k,i) is the frequency domain symbol corresponding to the i-th OFDM signal of the k-th subcarrier, and the PS filter at the transmitter is g k,i (n), g tx (n) is the pulse shape at the transmitting end, T is the duration of the discrete time domain OFDM signal, and n represents the nth sampling point in the discrete time domain; The channel is a multipath fading channel; The receiver includes a receiving end pulse shaping module, a coarse equalization module and a demodulation mapping module connected in sequence. The receiving end pulse shaping module includes a receiving DNN. The receiving DNN includes a receiving input layer for inputting the received signal and a receiving output layer. The receiving output layer is a receiving end PS filter. The received signal passes through the receiving end pulse shaping module to obtain a time domain received signal. In the coarse equalization module, the time domain received signal and the time domain channel are FFT-ed, and the two frequency domain information obtained are divided correspondingly. The division result is input into the demodulation mapping module. The demodulation mapping module is a neural network composed of multiple sub-neural networks. Each sub-neural network is implemented by DNN. The output of each sub-neural network is cascaded and spliced ​​to form the output of the demodulation mapping module.

2. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, characterized in that: The transmitting DNN also includes a transmitting hidden layer consisting of a layer of N neurons, a layer of 2N neurons and a power normalization layer, and the receiving DNN also includes a receiving hidden layer consisting of a layer of 2N neurons, a layer of N neurons and a power normalization layer.

3. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, characterized in that: The time domain received signal is Among them, y u To receive the signal, r k,i (n) is the PS filter at the receiving end, γ rx (n) is the pulse shape at the receiving end.

4. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, 2 or 3, characterized in that: In order to fit the ISI term in the received signal, at least two transmission signals are continuously transmitted into the channel, and the parameters and structures of the transmission DNN used by the at least two transmission signals are exactly the same.

5. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, 2 or 3, characterized in that: The modulation module is constellation modulation, so as to map the bit stream into frequency domain constellation symbols.

6. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, 2 or 3, characterized in that: The multipath fading channel is Where τ represents the maximum delay path number, δ(·) represents the discrete impulse response, and a p is the complex baseband channel coefficient of the pth delay path, τ p represents the delay of the pth delay path.

7. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, 2 or 3, characterized in that: The demodulation mapping module is composed of N / K sub-neural networks, and each sub-neural network is responsible for demodulating the signal on a subcarrier.

8. The CP-free PS-OFDM communication system based on an autoencoder according to claim 7, characterized in that: The demodulation mapping module includes an input layer connected to the coarse equalization module, an output layer formed by splicing the Softmax outputs of each sub-neural network, and a hidden layer. The hidden layer consists of a layer of 2K neurons, a layer of 4K neurons, a layer of 8K neurons, and a layer of 2K neurons.

9. The CP-free PS-OFDM communication system based on an autoencoder according to claim 8, characterized in that: The loss function of the demodulation mapping module is defined as in, corresponds to the output of the k-th Softmax classifier, indicating the probability of symbol demodulation on the k-th subcarrier, and M represents the modulation order of the modulation module, P k,1 ,...,P k,M Indicates the one-hot encoding used by the label, is the expected function.

10. The CP-free PS-OFDM communication system based on an autoencoder according to claim 1, 2 or 3, characterized in that: When training the transmitter, channel, and receiver, forward propagation training is used to calculate the loss function, and the gradient descent algorithm is used to complete back propagation and update the parameters until the training is completed.

Images