PAPR suppression algorithm realized based on pulse shaping modulation

Abstract

The invention provides a PAPR suppression algorithm realized based on pulse shaping modulation, and belongs to the field of communication, and the algorithm comprises the steps: building an end-to-end OFDM communication system model, the communication system model comprises a transmitter, a channel and a receiver, a transmitting end PS module is realized by a transmitting DNN, and the transmitting DNN comprises a transmitting input layer connected with an S / P module and a transmitting output layer; the channel is a multipath fading channel; the receiver comprises an equalization module and a demodulation mapping module which are connected in sequence; and S2, carrying out joint training on a neural network formed by the transmitter, the channel and the receiver by using a joint loss function L = alpha1L1 + alpha2L2. According to the invention, on the premise that the transmission rate and the error rate are not influenced, the PAPR of the transmitting end can be reduced, and flexible configuration of pulse shaping can be realized, so that different communication scenes can be flexibly supported.

Description

Technical Field

[0001] The invention relates to the field of communications, and in particular to a PAPR suppression algorithm based on pulse shaping modulation. Background Art

[0002] Orthogonal Frequency Division Multiplexing (OFDM) technology plays an important role in the fifth generation mobile communication system due to its robustness in multipath transmission and relatively low complexity of transceiver design. However, as a multi-carrier modulation technology, OFDM technology superimposes multiple orthogonal subcarriers, which will cause the envelope of the base station's transmitted signal to have large dynamic fluctuations, resulting in a high peak-to-average power ratio. High peak-to-average power ratio signals usually need to be transmitted through linear power amplifiers, but linear power amplifiers are very expensive and have higher process requirements, which will result in high hardware costs. If the high peak-to-average power ratio signal is passed through a nonlinear power amplifier, severe nonlinear distortion will occur, which will cause signal distortion and further lead to serious deterioration of system performance. The reserved subcarrier technology (Tone Reservation, TR) proposed by J.Tellado can be used to reduce the peak-to-average ratio (PAPR) of each antenna, but its PAPR suppression effect is not obvious and the data transmission rate will be significantly reduced; the PAPR suppression ability of the selective mapping method proposed by Jeon HB has been improved, but since each group of signals in the selective mapping method must be subjected to phase rotation, IFFT operation, calculation and comparison of PAPR values, the computational complexity is closely related to the number of candidate phases. The more candidate phases there are, the higher the computational complexity. In practical applications, it is necessary to comprehensively consider the PAPR suppression performance and computational complexity of the OFDM communication system to achieve a balance between the two.

[0003] Therefore, there is an urgent need to find a communication system that can effectively reduce PAPR without sacrificing transmission rate, control computational complexity, and avoid the defect of single pulse shape. Summary of the invention

[0004] The main purpose of the present invention is to propose a PAPR suppression algorithm based on pulse shaping modulation, which reduces the PAPR of the transmitter without affecting the transmission rate and bit error rate, and can 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 PAPR suppression algorithm based on pulse shaping modulation includes the following steps:

[0007] Step S1, build an end-to-end OFDM communication system model, the communication system model includes a transmitter, a channel and a receiver, the transmitter includes a constellation modulation module, an S / P module and a transmitter PS module connected in sequence, the transmitter PS module is implemented by a transmitting DNN, the transmitting DNN includes a transmitting input layer connected to the S / P module, and a transmitting output layer, the transmitting output layer is formed by N parallel PS filters, and the transmitting PS module outputs a time domain transmission signal s i =(G tx ⊙F H )x i ,s i is the main value sequence of the i-th time domain transmitted signal, x i represents the i-th frequency domain signal vector, is a stack of N PS filters corresponding to N subcarriers, F represents the Fourier transform matrix, (·) H represents conjugate transpose, ⊙ represents Hadamard product;

[0008] The channel is a multipath fading channel;

[0009] The receiver includes an equalization module and a demodulation mapping module connected in sequence. The equalization module performs FFT on the time domain received signal obtained through the channel and the channel in the time domain, and divides the two frequency domain information obtained by FFT by corresponding points to obtain the frequency domain information after equalization. The demodulation mapping module is implemented by a demodulation DNN. The demodulation DNN includes a demodulation input layer connected to the equalization module, and a demodulation output layer. The demodulation output layer is formed by splicing the Softmax outputs on N subcarriers to output the final demodulated bit value.

[0010] Step S2, jointly training the neural network composed of the transmitter, channel and receiver with a joint loss function L=α1L1+α2L2, where α1 represents the ratio of the transmitter loss function, α2 represents the ratio of the receiver loss function, s(n) is s i The nth element of represents the probability of symbol demodulation on the kth subcarrier, and M is the modulation order, P k,1 ,...,P k,M Indicates the unique hot encoding used by the label, and its value range is {0, 1}, is the expected function.

[0011] Furthermore, the transmitting DNN also includes a transmitting hidden layer consisting of a layer of N neurons, a first Dropout layer, a layer of 2N neurons, a second Dropout layer, a layer of 2N neurons and a power normalization layer.

[0012] Furthermore, the demodulation DNN also includes a demodulation hidden layer consisting of a layer of 2N neurons, a layer of 4N neurons, a layer of 8N neurons and a layer of 2N neurons, and the first layer and the last layer of the demodulation hidden layer use jump connections.

[0013] Furthermore, the input data of the transmitting DNN and the demodulating DNN need to be preprocessed, and the preprocessing includes splitting the real part and the imaginary part of each original data, sorting the real part of each original data to form a real data block, and sorting the imaginary part of each original data to form an imaginary data block.

[0014] Furthermore, in step S1, the transmitting end PS module obtains the signal s i =(G tx ⊙F H )x i Finally, a cyclic prefix is ​​added to the signal to combat frequency selective fading.

[0015] Furthermore, the neuron ratio parameter of the first Dropout layer and the second Dropout layer is set to 0.5.

[0016] Furthermore, the step S2 specifically includes:

[0017] Step S21: Get 10 6 The group data is used as a dataset, the dataset is divided into a training set and a validation set, the joint loss function L = α1L1 + α2L2 is selected, and the hyperparameters are set;

[0018] Step S22: forward propagation training, calculating the joint loss function, and selecting a suitable gradient descent algorithm to complete back propagation and parameter update until the training is completed;

[0019] Step S23, statistically calculate the loss curve and the accuracy fitting curve, and pay attention to the training time. If it does not converge, adjust the hyperparameters and return to step S21. If it converges, go to step S24;

[0020] Step S24, obtain 10 5 The group data is used as a test set, and the test set is input to test the performance by statistically analyzing the PAPR and BER in the test set.

[0021] Furthermore, in the joint loss function, α1:α2=0.01:3.

[0022] Furthermore, the multipath fading channel is expressed as 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.

[0023] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The transmitting DNN of the transmitting PS module of the present invention, the channel and the demodulation DNN of the receiving end constitute an encoder, and a sophisticated pulse shaping mechanism is introduced within the framework of the traditional OFDM system, so as to optimize the transmission waveform and reduce the PAPR of the transmitting signal. The transmitting DNN and the demodulation DNN are jointly trained with a joint loss function to simultaneously achieve the purpose of suppressing the PAPR of the transmitting signal and preventing the deterioration of the system bit error rate, thereby achieving a balance between reducing the PAPR and controlling the calculation complexity without affecting the transmission rate and the bit error rate, and realizing flexible configuration of pulse shaping to avoid the defect of a single pulse shape, so as to flexibly support different communication scenarios.

[0025] 2. The present invention preprocesses the input data of the transmitting DNN and the demodulating DNN, which can achieve better performance without affecting the transmission rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0027] Figure 1 It is a principle block diagram of the end-to-end OFDM communication system model of the present invention.

[0028] Figure 2 Schematic diagram of the PS filter of the present invention.

[0029] Figure 3 Schematic diagram of the structure of the launch DNN of the present invention.

[0030] Figure 4 This is a schematic diagram of data preprocessing of the present invention.

[0031] Figure 5 This is a schematic diagram before and after using the first dropout layer or the second dropout layer in the present invention.

[0032] Figure 6 It is a loss function curve diagram of the training set and the validation set during the training process of the present invention.

[0033] Figure 7 The PAPR performance comparison diagram of the present invention and the CP-OFDM system under different training hyperparameters when 16QAM modulation is adopted.

[0034] Figure 8 This is a comparison chart of the BER performance of the present invention and the CP-OFDM system under different training hyperparameters when 16QAM modulation is adopted. DETAILED DESCRIPTION

[0035] The present invention is further described below through specific implementation modes.

[0036] The PAPR suppression algorithm based on pulse shaping modulation includes the following steps:

[0037] Step S1: Build Figure 1 The end-to-end OFDM communication system model shown in the figure includes a transmitter, a channel and a receiver. The function of the transmitter is to transform the bit stream into an actual time domain transmission signal, including a constellation modulation module, an S / P module, a transmitter PS (pulse shaping) module and a CP module connected in sequence. The constellation modulation module maps the bit stream into a frequency domain constellation symbol according to the selected modulation mode. When the QPSK modulation mode is selected, the constellation modulation module converts continuous 2-bit bit information into 1 frequency domain symbol. The frequency domain symbol value range is {0.71+0.71j, 0.71-0.71j, -0.71+0.71j, -0.71-0.71j}. The S / P module is used to convert the serial frequency domain signal into a parallel frequency domain signal.

[0038] The transmitting end PS module is used to adjust the power distribution of the transmitting signal in the time-frequency domain, thereby optimizing the PAPR. It is implemented by the transmitting DNN. The transmitting DNN includes a transmitting input layer, a transmitting hidden layer, and a transmitting output layer connected to the output end of the S / P module. The input of the transmitting input layer is an OFDM time domain signal. The transmitting output layer is formed by N parallel PS filters. The transmitting end PS module outputs a time domain transmitting signal. 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 corresponding to the i-th subcarrier at the transmitter is 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. In this embodiment, the length of the PS filter is limited to N, that is, T = N. The time domain transmission signal can be rewritten as a matrix form s i =(G tx ⊙F H )x i ,s i is the main value sequence of the i-th time domain transmitted signal, x i represents the i-th frequency domain signal vector, is a stack of N PS filters corresponding to N subcarriers, F represents the Fourier transform matrix, (·) H represents the conjugate transpose, ⊙ represents the Hadamard product, such as Figure 2 Shown is the schematic diagram of the PS filter, Figure 2 middle, is the filter on the kth subcarrier The nth value of .

[0039] The PS module at the transmitting end receives the signal s i =(G tx ⊙F H )x i Finally, a cyclic prefix (CP) of corresponding length is added to the signal in the CP module to combat frequency selective fading.

[0040] The structure of the launch DNN is as follows Figure 3 As shown in , the emission hidden layer consists of a layer of N neurons, a first Dropout layer, a layer of 2N neurons, a second Dropout layer, a layer of 2N neurons, and a power normalization layer. The neuron ratio parameter of the first Dropout layer and the second Dropout layer is set to 0.5, indicating that 50% of the neurons will be inactivated in this layer. There are no trainable parameters in the power normalization layer. Figure 5 Shown is a comparison diagram before and after using the Dropout layer.

[0041] The PS module at the transmitter is actually designed with the phase value on the PPN (Poly-Phase Network). The data is input into the neural network to obtain the specific required phase value, and then the PPN implements the specific function, providing a method of digital design and analog device implementation.

[0042] The channel is a multipath fading channel, expressed as 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 Denotes the delay of the pth delay path. The multipath fading channel can be regarded as a channel layer with no optimized parameters. There are no internal training parameters, and it only realizes the conversion between the transmitted signal and the received signal.

[0043] For this channel, the conversion between the transmitted signal x and the received signal y can be expressed as: u =ΔHx u ++w u

[0044] Among them, x u Represents the time domain transmitted signal at the current moment, that is, for the received signal y at the current moment u For a wireless receiver, due to the influence of multipath and CP, the received signal is the circular convolution of the channel vector and the transmitted signal.

[0045] The receiver does not perform pulse shaping optimization, so the receiver processes the received signal after removing the CP, including an equalization module and a demodulation mapping module connected in sequence. The equalization module performs N-point FFT on the time domain received signal and the time domain channel obtained through the channel, and performs corresponding point division on the two N-point frequency domain information sequences obtained by FFT to obtain the equalized frequency domain information, which is input into the demodulation mapping module together with the frequency domain channel information. The demodulation mapping module is implemented by a demodulation DNN, which includes a demodulation input layer, a demodulation hidden layer and a demodulation output layer connected to the equalization module. The demodulation output layer is composed of the Softmax output on N subcarriers, and the dimension is consistent with the number of information bits to output the final demodulated bit value. The demodulation hidden layer consists of a layer of 2N neurons, a layer of 4N neurons, a layer of 8N neurons and a layer of 2N neurons. The first and last layers of the demodulation hidden layer use jump connections to prevent the gradient from disappearing.

[0046] Step S2, jointly training the neural network composed of the transmitter, channel and receiver with a joint loss function L=α1L1+α2L2, where α1 represents the ratio of the transmitter loss function, α2 represents the ratio of the receiver loss function, x(n) is s i The nth element of corresponds to the output of a Softmax classifier, indicating the probability of symbol demodulation on the kth subcarrier, and M is the modulation order, P k,1 ,...,P k,M Indicates the unique hot encoding used by the label. Since the output purpose is to directly restore the transmission bit quality, P k,1 ,...,P k,M The value range is {0, 1}, is the expected function;

[0047] The specific steps include:

[0048] Step S21: Get 10 6 The group data is used as a data set, and the data set is divided into a training set and a validation set (the training set and the validation set are divided according to 8:2). The labels and data sets are input, the joint loss function L = α1L1 + α2L2 is selected, and the hyperparameters are set. The hyperparameters are shown in Table 1:

[0049] Table 1

[0050] Hyperparameters illustrate N Number of subcarriers Drop_rate Dropout neuron ratio <![CDATA[α1]]> Transmitter loss function ratio <![CDATA[α2]]> Receiver loss function ratio lr Initial learning rate batch_size Mini-batch values dl Learning rate decay exponential

[0051] The neural network is built based on Tensorflow 2.1 and Kreas 2.3.1 environments. For the joint loss function L = α1L1 + α2L2, if α1 accounts for a large proportion, the neural network will pay more attention to improving the PAPR suppression capability of the OFDM system during training; if α1 accounts for a small proportion, the neural network will pay more attention to reducing the bit error rate of the system during training.

[0052] Step S22: forward propagation training, calculating the joint loss function, and selecting a suitable gradient descent algorithm to complete back propagation and parameter update until the training is completed;

[0053] 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;

[0054] Step S23, statistically calculate the loss curve and the accuracy fitting curve, and pay attention to the training time. If it does not converge, adjust the hyperparameters and return to step S21. If it converges, go to step S24;

[0055] 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;

[0056] Step S24, obtain 10 5 The group data is used as a test set, and the test set is input to test the performance by statistically analyzing the PAPR and BER in the test set.

[0057] In order to ensure the quality of data and prevent the loss of imaginary information, and further improve the performance of neural networks, the input data of the transmitting DNN and the demodulating DNN need to be preprocessed, such as Figure 4 As shown, the preprocessing includes splitting the real part and the imaginary part of each original data, and sorting and combining the real part of each original data to form a real data block, and sorting and combining the imaginary part of each original data to form an imaginary data block.

[0058] Figure 6 This is the loss function curve when α1=0.1,α2=10. Figure 5 The trend of the loss function of the training set and the validation set during training is intuitively depicted in Figure 2. By observing this curve, we can see that the performance of the neural network on the training set and the validation set is almost the same. The neural network shows good stability and reliability when predicting unknown data, which means that the knowledge learned by the neural network on the training set is also applicable to the validation set, thus proving that the neural network has good generalization ability.

[0059] Figure 7 and Figure 8The performance changes of the neural network under different loss function weight ratios α1:α2 are shown: when α1 accounts for a small proportion, the neural network pays more attention to improving the system's bit error rate performance; when α1 accounts for a large proportion, the neural network pays more attention to improving the system's PAPR suppression ability. Combined with the above data, the present invention finally chooses to complete the training of the neural network when the loss function weight ratio is α1:α2=0.01:3, and obtains the phase value required to achieve the PPN of the analog device. It can be seen that the PAPR suppression ability of the system at this time is relatively good compared to the CP-OFDM system ( Figure 7 and Figure 8 The curve corresponding to the "original signal" in the figure has been significantly improved, and the BER performance has no obvious loss, and does not affect the transmission rate of the system. Figure 7 The ordinate is the complementary cumulative distribution function (CCDF), which is used to represent the statistical characteristics of the peak-to-average PAPR in the OFDM system.

[0060] In the present invention, the terms "first", "second", "third", etc. are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. In the description, the directions or positional relationships indicated by "upper", "lower", "left", "right", "front", and "back" are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention, and are not intended to indicate or imply that the device referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the scope of protection of the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0061] In addition, in the description of this application, unless otherwise specified, "plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0062] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A PAPR suppression algorithm based on pulse shaping modulation, characterized in that: The steps include: Step S1, build an end-to-end OFDM communication system model, the communication system model includes a transmitter, a channel and a receiver, the transmitter includes a constellation modulation module, an S / P module and a transmitter PS module connected in sequence, the transmitter PS module is implemented by a transmitting DNN, the transmitting DNN includes a transmitting input layer connected to the S / P module, and a transmitting output layer, the transmitting output layer is formed by N parallel PS filters, and the transmitting PS module outputs a time domain transmission signal s i =(G tx ⊙F H )x i ,s i is the main value sequence of the i-th time domain transmitted signal, x i represents the i-th frequency domain signal vector, is a stack of N PS filters corresponding to N subcarriers, F represents the Fourier transform matrix, (·) H represents conjugate transpose, ⊙ represents Hadamard product; The channel is a multipath fading channel; The receiver includes an equalization module and a demodulation mapping module connected in sequence. The equalization module performs FFT on the time domain received signal obtained through the channel and the channel in the time domain, and divides the two frequency domain information obtained by FFT by corresponding points to obtain the frequency domain information after equalization. The demodulation mapping module is implemented by a demodulation DNN. The demodulation DNN includes a demodulation input layer connected to the equalization module, and a demodulation output layer. The demodulation output layer is formed by splicing the Softmax outputs on N subcarriers to output the final demodulated bit value. Step S2, jointly training the neural network composed of the transmitter, channel and receiver with a joint loss function L=α1L1+α2L2, where α1 represents the ratio of the transmitter loss function, α2 represents the ratio of the receiver loss function, s(n) is s i The nth element of represents the probability of symbol demodulation on the kth subcarrier, and M is the modulation order, P k,1 ,...,P k,M Indicates the unique hot encoding used by the label, and its value range is {0, 1}, is the expected function.

2. The PAPR suppression algorithm based on pulse shaping modulation according to claim 1, characterized in that: The transmitting DNN also includes a transmitting hidden layer consisting of a layer of N neurons, a first Dropout layer, a layer of 2N neurons, a second Dropout layer, a layer of 2N neurons and a power normalization layer.

3. The PAPR suppression algorithm based on pulse shaping modulation according to claim 2, characterized in that: The demodulation DNN also includes a demodulation hidden layer consisting of a layer of 2N neurons, a layer of 4N neurons, a layer of 8N neurons and a layer of 2N neurons, and the first layer and the last layer of the demodulation hidden layer use jump connections.

4. A PAPR suppression algorithm based on pulse shaping modulation according to claim 1, 2 or 3, characterized in that: The input data of the transmitting DNN and the demodulating DNN need to be preprocessed, and the preprocessing includes splitting the real part and the imaginary part of each original data, sorting the real part of each original data to form a real data block, and sorting the imaginary part of each original data to form an imaginary data block.

5. A PAPR suppression algorithm based on pulse shaping modulation according to claim 1, 2 or 3, characterized in that: In step S1, the transmitting end PS module obtains the signal s i =(G tx ⊙F H )x i Finally, a cyclic prefix is ​​added to the signal to combat frequency selective fading.

6. A PAPR suppression algorithm based on pulse shaping modulation according to claim 2 or 3, characterized in that: The neuron ratio parameter of the first Dropout layer and the second Dropout layer is set to 0.

5.

7. A PAPR suppression algorithm based on pulse shaping modulation according to claim 1, 2 or 3, characterized in that: The step S2 specifically includes: Step S21: Get 10 6 The group data is used as a dataset, the dataset is divided into a training set and a validation set, the joint loss function L = α1L1 + α2L2 is selected, and the hyperparameters are set; Step S22: forward propagation training, calculating the joint loss function, and selecting a suitable gradient descent algorithm to complete back propagation and parameter update until the training is completed; Step S23, statistically calculate the loss curve and the accuracy fitting curve, and pay attention to the training time. If it does not converge, adjust the hyperparameters and return to step S21. If it converges, go to step S24; Step S24, obtain 10 5 The group data is used as a test set, and the test set is input to test the performance by statistically analyzing the PAPR and BER in the test set.

8. The PAPR suppression algorithm based on pulse shaping modulation according to claim 6, characterized in that: In the joint loss function, α1:α2=0.01:

3.

9. A PAPR suppression algorithm based on pulse shaping modulation according to claim 1, 2 or 3, characterized in that: The multipath fading channel is expressed as 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.

Images