A distribution optimization method for probability shaping models in turbulent channels

Abstract

The present invention discloses a probability shaping model distribution optimization method suitable for turbulent channels, which is based on neural networks and genetic algorithms, wherein the neural network is used to predict the generalized mutual information (GMI) about the input distribution, and the genetic algorithm is used to select the optimal input distribution corresponding to the current turbulent channel based on the trained neural network. The turbulent channel is different from the Gaussian white noise (AWGN) channel and belongs to a nonlinear channel. The probability shaping signal distribution technology adopted in the nonlinear channel can improve the system error performance. The probability shaping signal can achieve different source entropies based on different distribution parameters or different truncation schemes. The present invention is aimed at turbulent channels, realizes the selection of the most suitable shaping distribution corresponding to turbulent channel conditions with different nonlinear intensities, further improves the signal's tolerance and power limit for the nonlinearity of the turbulent channel, reduces the bit error rate, and improves the information rate of the system, thereby maximizing the system transmission capacity.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a probability shaping model distribution optimization method suitable for turbulent channels. Background Art

[0002] In recent years, with the development of technologies such as the Internet of Things, cloud computing, virtual reality (VR) and augmented reality (AR), people have an increasing demand for communication transmission rates. Unlike optical fiber communications and traditional radio communications, free space optics (FSO) communications have the characteristics of high bandwidth, easy deployment, low power consumption, low quality and high security, which can complement other communication methods well. At the same time, adding new elements to traditional FSO communications, such as probabilistic shaping, can improve its relatively weak transmission quality.

[0003] Atmospheric turbulence and attenuation effects are the main considerations for the environment in which the FSO channel is located. When the optical signal is transmitted in it, it will be attenuated by the absorption and scattering effects of particles in the atmosphere, and will interact with turbulence to produce complex effects. The turbulence effect corresponding to atmospheric turbulence refers to the random change of the refractive index of turbulent elements at different spatial positions in the atmosphere, which leads to light intensity flickering and phase fluctuations. It is a nonlinear disturbance, and the intensity of the optical signal obtained by the receiver of the FSO system will fluctuate. For space optical communication systems affected by turbulence, the higher the power constellation point in the constellation diagram, the more obvious the impact of turbulence, the more serious the distortion, the more likely it is to produce bit errors, and reduce the communication quality of the system.

[0004] The application of probabilistic shaping technology in fiber channel transmission has proven that it has a good effect on expanding the capacity of nonlinear channel systems and solving power limitation problems. Therefore, applying the probabilistic shaping technology based on distributed parameters to FSO channels based on nonlinear turbulence phenomena is a feasible method with the potential to increase capacity.

[0005] Neural networks have been gradually applied to communication systems. However, for unknown nonlinear functions, it is difficult to accurately find the function extreme value only through the input and output data of the function. Therefore, the present invention uses a neural network (NN) combined with a genetic algorithm (GA) to solve, and uses the nonlinear fitting ability of the neural network and the nonlinear optimization ability of the genetic algorithm to find the function extreme value. According to the characteristics of the channel, the probability of occurrence of different power constellation points can be optimally adjusted to achieve the maximum achievable transmission rate (AIR), so it is particularly suitable for space optical communication scenarios based on turbulence. Summary of the invention

[0006] The purpose of the present invention is to provide a probability shaping model distribution optimization method suitable for turbulent channels, by finding the optimal distribution corresponding to different turbulent channel environments, so as to reduce the system error performance and improve the transmission rate.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] In one aspect, the present invention provides a probability shaping model distribution optimization method applicable to a turbulent channel, comprising the following steps:

[0009] S1. In the signal transmission module, different source entropies are achieved based on different distribution parameters or different truncation schemes;

[0010] S2, after the signal is generated from the transmitting module, it is transmitted through the nonlinear turbulent channel and reaches the signal receiving module. In the signal receiving module, the signal after down-conversion and resampling is first digitally processed, and then the GMI of the signal receiving module is calculated;

[0011] S3. Under the current turbulent channel intensity, a neural network is used to approximate a nonlinear multivariable function based on the distribution obtained by the signal transmitting module and the corresponding generalized mutual information of the signal receiving module. The multivariable function maps the input distribution D of the signal transmitting module to the GMI of the signal receiving module as the performance measure of the shaping scheme;

[0012] S4. The trained neural network is used as a multivariable function fNN, and the genetic algorithm is used to optimize the input distribution to select the optimal input distribution corresponding to the current turbulent channel.

[0013] Furthermore, in step S1, the process of the signal transmission module generating a Truncate-SGPS-m-QAM shaped signal is as follows: the distribution parameters include the modulation order m, the scaling factor v, the super Gaussian factor s and the truncation threshold t, wherein the scaling factor and the super Gaussian factor are based on the Maxwell-Boltzmann distribution, and the probability distribution of all QAM symbols is calculated through a single-path probability distribution, and then the truncation threshold t is set based on the probability mass function of the two-dimensional QAM signal, and the symbols with PMF lower than t are deleted to obtain the Truncate-SGPS-m-QAM shaped signal.

[0014] Furthermore, in step S2, the process of digital signal processing performed by the signal receiving module includes: orthogonal normalization, clock recovery, channel equalization, frequency offset estimation and phase recovery.

[0015] Furthermore, the training method of the neural network in step S3 is: constructing a BP neural network, using the Log-Sigmoid function as the activation function of the hidden layer, using the ReLU activation function in the output layer, forward propagating the working signal, and backward propagating the error signal, and continuously adjusting the neuron weights to minimize the mean square error between the output of the neural network and the GMI from the training set; extracting the distribution of the signal sending module (D1, D2, D3...) and the generalized mutual information of the signal receiving module (GMI1, GMI2, GMI3...) to form a data sample set, in which the training data accounts for 75% and the test data accounts for 25%.

[0016] Furthermore, the genetic algorithm in step S4 searches for an approximate optimal solution by setting fNN as a fitness function and performing mutation, crossover and selection on the input part of the code.

[0017] On the other hand, the present invention also provides the application of the above-mentioned probability shaping model distribution optimization method suitable for turbulent channels in a space optical communication scenario based on turbulent channels.

[0018] Among them, the current changes in the turbulent channel include changes in the transmission distance, the optical power at the transmitting end, and the transmission rate.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The present invention proposes a probability shaping model distribution optimization method suitable for turbulent channels, which is based on neural networks and genetic algorithms, wherein the neural network is used to predict the generalized mutual information (GMI) about the input distribution, and the genetic algorithm is used to select the optimal input distribution corresponding to the current turbulent channel based on the trained neural network. The turbulent channel is different from the Gaussian white noise (AWGN) channel and belongs to a nonlinear channel. The use of probability shaping signal distribution technology in the nonlinear channel can improve the system error performance. The probability shaping signal can achieve different source entropies based on different distribution parameters or different truncation schemes. The present invention is aimed at turbulent channels and realizes the selection of the most suitable shaping distribution corresponding to turbulent channel conditions with different nonlinear intensities, further improving the signal's tolerance and power limit for the nonlinearity of the turbulent channel, reducing the bit error rate, and improving the information rate of the system, thereby maximizing the system transmission capacity and achieving the maximum achievable transmission rate. It is particularly suitable for space optical communication scenarios based on turbulent channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention, and for those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0022] Figure 1 A principle block diagram of a probability shaping model distribution optimization method applicable to a turbulent channel provided in an embodiment of the present invention.

[0023] Figure 2 Schematic diagram of a Truncate-SGPS-m-QAM signal generated by a signal transmission module provided in an embodiment of the present invention.

[0024] Figure 3 The constellation diagram of probability shaping (PS) and super-Gaussian distribution probability shaping (SGPS) provided in the embodiment of the present invention under AWGN channel (30dB) and the PMF distribution bar graph of their I / Q single channels respectively.

[0025] Figure 4 A schematic diagram of a Truncate-SGPS-m-QAM signal generated by truncating a Super Gaussian Distribution Probability Shaping (SGPS) signal provided in an embodiment of the present invention.

[0026] Figure 5 A schematic diagram of a neural network with multiple hidden layers for predicting the transmission end distribution and the receiving end GMI provided by an embodiment of the present invention.

[0027] Figure 6 A flow chart of a genetic algorithm provided by an embodiment of the present invention.

[0028] Figure 7 A constellation diagram of the receiving end of the Truncate-SGPS-m-QAM signal provided in an embodiment of the present invention after transmission in a spatial optical channel affected by turbulence, as well as a comparison diagram of the transmission results with traditional uniformly distributed signals, PS signals, SGPS signals and Truncate-PS signals. DETAILED DESCRIPTION

[0029] In order to better understand the technical solution, the method of the present invention is described in detail below with reference to the accompanying drawings.

[0030] The probability shaping model distribution optimization method of the present invention is suitable for turbulent channels, such as Figure 1 As shown, the following steps are included:

[0031] 1. In the signal transmission module, different source entropies can be achieved based on different distribution parameters or different truncation schemes

[0032] Taking the m-QAM signal as an example, a Truncate-SGPS-m-QAM shaped signal can be generated in the transmitting module, where Truncate means truncation in Chinese, which represents selecting a threshold of the truncation probability according to the constellation point mass distribution function (PMF). SGPS (Super-Gaussian probabilistic shaping) means super-Gaussian probability shaping in Chinese, which represents changing the signal constellation point distribution by changing the super-Gaussian factor and the scaling factor. m-QAM stands for m-order quadrature amplitude modulation.

[0033] Specific as Figure 2 As shown, by designing the distribution parameters, Truncate-SGPS-m-QAM shaped signals with different distributions are generated in the transmitting module, that is, truncated super-Gaussian probability shaped signals.

[0034] The distribution parameters of the signal include: modulation order m, scaling factor v, super Gaussian factor s, and truncation threshold t, wherein the scaling factor and super Gaussian factor are based on the Maxwell-Boltzmann distribution, as shown in the following formula:

[0035]

[0036] Among them, P x Represents each constellation point x in the constellation diagram i The corresponding occurrence probability is 1≤i≤m.

[0037] This is only the calculation formula for one-dimensional symbols, that is, the I or Q path of the QAM signal. The probability distribution of all QAM signals can be calculated through the probability distribution of a single path. Then, based on the probability mass function (PMF) of the two-dimensional QAM signal, the truncation threshold t is set, and the symbols with PMF lower than t are deleted to obtain the Truncate-SGPS-m-QAM shaped signal.

[0038] The specific steps are as follows:

[0039] 1. First, two Super Gaussian (SG) PS pulse amplitude modulation (PAM) signals are generated. A constant constant composition distribution matcher (CCDM) is usually used to convert the uniformly distributed input bits into the positive amplitude of the SGPS-PAM signal (a "half PAM" constellation); the constellation diagram of probability shaping (PS) and super Gaussian distribution probability shaping (SGPS) in the AWGN channel (30dB) and the PMF distribution histogram of the I / Q single channel are shown as follows Figure 3 As shown;

[0040] 2. Create SGPS-PAM with the desired probability distribution by using the parity bits generated by the system forward error correction (FEC) encoder as sign bits;

[0041] 3. Combine two orthogonal SGPS-PAM signals to generate SGPS-QAM signal;

[0042] 4. Such as Figure 4 As shown, according to a pre-set truncation threshold, the constellation points in the SGPS-QAM symbol whose PMF is less than the truncation threshold are deleted to obtain a Truncate-SGPS-m-QAM signal.

[0043] 2. The optical signal after electro-optical modulation is transmitted through the nonlinear turbulent channel and reaches the signal receiving module

[0044] When the wireless optical signal passes through the turbulent channel and enters the receiving module, it is first down-converted into an electrical baseband signal and then resampled. Figure 1 The DSP (digital signal processing) module shown in the figure is used for processing, where the DSP processing part should include: orthogonal normalization, clock recovery, channel equalization, frequency offset estimation, and phase recovery. The specific functions implemented are:

[0045] 1. Orthogonal normalization: It is used to solve the problem of inconsistent extinction ratios of the two arms due to the process problems of the IQ modulator, or the different driving amplitudes of the driving signals of the two arms due to the process problems of the electrical amplifier, or the inconsistent responses of the optical mixer or balanced detector at the receiver end, which leads to the non-orthogonality or imbalance of the two IQ paths.

[0046] 2. Clock recovery: used to solve the system sampling error caused by the local sampling clock not being synchronized with the transmitter signal clock or the instability of the local clock source itself.

[0047] 3. Channel equalization: For polarization multiplexing systems, since optical signals are disturbed by the environment in turbulent channels, their polarization mode dispersion is constantly changing, so we need to dynamically equalize the channel.

[0048] 4. Frequency deviation estimation: When using coherent light reception, there is a certain amount of frequency deviation between the local oscillator light of the receiving module and the signal light of the transmitting module. At the same time, as the communication time increases, frequency drift, that is, frequency deviation, is inevitable between the lasers at both ends of the transmission and reception. Therefore, digital signal processing algorithms are used to compensate for the impact of frequency deviation on the signal.

[0049] 5. Phase recovery: In QAM modulation systems, frequency deviation is usually converted into phase offset and compensated using digital signal processing algorithms.

[0050] Perform GMI calculation on the signal processed by DSP to obtain the GMI of the corresponding transmitting end distribution under the channel environment. The calculation formula of GMI is:

[0051]

[0052] Among them, y k is the kth symbol of the received noise signal sequence of length n, b k,i is the i-th bit of the k-th transmitted symbol, and is the i-th bit in the QAM constellation point is b k,i The symbol collection, q Y|X It is the conditional probability based on the channel, specifically:

[0053]

[0054] Among them, x is the transmitted signal, y is the received signal, and N0 is the channel noise variance.

[0055] 3. Repeatedly change the distribution parameters of the transmitting signal to obtain the corresponding new GMI and construct a BP neural network.

[0056] like Figure 5 As shown, a multi-hidden layer neural network with 4 inputs (m, v, s, t) and 1 output (GMI) is constructed according to the data set to predict the generalized mutual information GMI. The above steps 1 and 2 are repeated to extract the transmission end distribution parameters (D1, D2, D3...) and the receiving end generalized mutual information (GMI1, GMI2, GMI3...) to form a data sample set. The distribution parameters D include (m, v, s, t), where m is the QAM signal order, v is the scaling factor, s is the super Gaussian factor, and t is the truncation threshold. The training data accounts for 75% and the test data accounts for 25%.

[0057] The constructed BP neural network uses the Log-Sigmoid function as the activation function of the hidden layer as follows:

[0058]

[0059] The output layer uses the ReLU activation function, the working signal is forward propagated, the error signal is back propagated, and the neuron weights are continuously adjusted to minimize the mean square error (MSE) between the output of the neural network and the GMI from the training set.

[0060] 4. Find the distribution that maximizes the function output

[0061] The trained neural network, as a multivariable function fNN, finds the distribution that maximizes the function output, which is equivalent to an optimization problem. Since this is a non-convex nonlinear minimization problem, we use a genetic algorithm to find an approximate optimal solution by setting fNN as the fitness function and mutating, crossing, and selecting the encoded input part.

[0062] The multivariable function fNN is input into a genetic algorithm (GA) module, such as Figure 6As shown in the figure, the population is initialized first, and a batch of D(m,v,s,t) combinations are randomly generated, representing multiple probability distribution models using different modulation formats, shaping coefficients, super-Gaussian orders, and stage thresholds. Among them, each group of D(m,v,s,t) combinations is called an individual, and the set of all individuals together is called a population. The values ​​of the four parameters m, v, s, and t in each individual are called the individual genes. Then, the fitness corresponding to each individual is calculated according to the multivariate function fNN, that is, the GMI value predicted by fNN. After that, a selection operation is performed to select individuals with high fitness from the current population. Then, a crossover operation is performed to randomly pair different individuals in the screened population, and the values ​​of some genes are exchanged after pairing. The next step is to perform a mutation operation to mutate some genes in the new individuals generated after the crossover. Finally, the process from fitness calculation to mutation is repeated until the maximum number of iterations is reached or the fitness value has converged. After the iteration, the global optimal value corresponding to fNN and the corresponding input value, that is, the optimal GMI and its corresponding input distribution D, can be obtained. This module optimizes the input distribution of the desired spectral efficiency to find the optimal distribution for the current turbulent channel environment.

[0063] like Figure 7 As shown, in a spatial optical transmission system affected by turbulence, when the Gamma-Gamma turbulence model is used to simulate the channel characteristics, considering that the system Rytov variance is 0.05 and the signal-to-noise ratio of the additive white Gaussian noise is 19dB, different signal probability distribution models are used, and the clarity of the constellation diagram at the receiving end and the achievable GMI performance have obvious differences. Among them, using the traditional uniform distribution, the transmission performance is the worst, and the constellation points in the outer circle of the constellation diagram cannot be separated due to the influence of turbulence. After using the PS technology, the probability of occurrence of the outer circle constellation points is reduced, the influence of turbulence is alleviated, and the GMI performance is improved to a certain extent, but the global optimum is not achieved. Using SGPS and Truncate-PS technology alone can further improve the GMI performance of the spatial optical transmission system to a small extent on the basis of PS. Finally, by using the probability shaping model distribution optimization method proposed in the present invention, the Truncate-SGPS model matching the channel characteristics is used for communication, and the optimal GMI performance can be achieved.

[0064] Based on the above principles, the present invention proposes a probability shaping model distribution optimization method suitable for turbulent channels. By constructing a neural network and using a genetic algorithm to optimize the extreme values ​​of nonlinear functions, the most suitable shaping distribution is selected corresponding to turbulent channel conditions with different nonlinear intensities. The flexibility of the shaping technology is fully utilized, the system error performance and the system information rate are improved, thereby maximizing the transmission capacity of the turbulent channel system.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A probability shaping model distribution optimization method suitable for turbulent channels, characterized in that: The following steps are involved: S1. In a signal transmission module, different source entropies are realized based on different distribution parameters or different truncation schemes; the signal transmission module generates truncated super-Gaussian probability shaping signals Truncate-SGPS-m-QAM shaping signals with different distributions; the process of the signal transmission module generating the Truncate-SGPS-m-QAM shaping signal is as follows: the distribution parameters include the modulation order m, the scaling factor v, the super-Gaussian factor s and the truncation threshold t, wherein the scaling factor and the super-Gaussian factor are based on the Maxwell-Boltzmann distribution, and the probability distribution of all QAM symbols is calculated through the probability distribution of a single channel, and then the truncation threshold t is set based on the probability mass function of the two-dimensional QAM signal, and the symbols with PMF lower than t are deleted to obtain the Truncate-SGPS-m-QAM shaping signal; S2, after the signal is generated from the transmitting module, it is transmitted through the nonlinear turbulent channel and reaches the signal receiving module. In the signal receiving module, the signal after down-conversion and resampling is first digitally processed, and then the GMI of the signal receiving module is calculated; S3. Under the current turbulent channel intensity, a neural network is used to approximate a nonlinear multivariable function based on the distribution obtained by the signal transmitting module and the corresponding generalized mutual information of the signal receiving module. The multivariable function maps the input distribution D of the signal transmitting module to the GMI of the signal receiving module as the performance measure of the shaping scheme; S4. The trained neural network is used as a multivariable function fNN, and the genetic algorithm is used to optimize the input distribution to select the optimal input distribution corresponding to the current turbulent channel.

2. The probability shaping model distribution optimization method suitable for turbulent channels according to claim 1 is characterized in that: Step S2: The process of digital signal processing by the signal receiving module includes: orthogonal normalization, clock recovery, channel equalization, frequency offset estimation and phase recovery.

3. The probability shaping model distribution optimization method suitable for turbulent channels according to claim 1 is characterized in that: The training method of the neural network in step S3 is as follows: construct a BP neural network, use the Log-Sigmoid function as the activation function of the hidden layer, use the ReLU activation function in the output layer, forward propagate the working signal, and reverse propagate the error signal, and continuously adjust the neuron weights to minimize the mean square error between the output of the neural network and the GMI from the training set; extract the distribution of the signal sending module (D1, D2, D3...) and the generalized mutual information of the signal receiving module (GMI1, GMI2, GMI3...) to form a data sample set, in which the training data accounts for 75% and the test data accounts for 25%.

4. The probability shaping model distribution optimization method suitable for turbulent channels according to claim 1 is characterized in that: The genetic algorithm in step S4 searches for an approximate optimal solution by setting fNN as a fitness function and performing mutation, crossover and selection on the input part of the code.

5. Application of the probability shaping model distribution optimization method suitable for turbulent channels according to any one of claims 1 to 4 in a space optical communication scenario based on turbulent channels.

6. The probability shaping model distribution optimization method suitable for turbulent channels according to claim 1, characterized in that: The current changes in the turbulent channel include changes in the transmission distance, the optical power at the transmitter, and the transmission rate.

Images