CN-122001474-A - End-to-end optimization method, device and system for probability shaping signal

CN122001474ACN 122001474 ACN122001474 ACN 122001474ACN-122001474-A

Abstract

The invention provides a probability shaping signal end-to-end optimization method, device and system, wherein the method comprises the steps of receiving a disturbed signal from a sending end and transmitting and outputting the disturbed signal through a fiber channel in each iteration period; and performing end-to-end joint training on the probability generator and the decoder through a joint transmitting end so that the transmitting end can obtain a target probability shaping signal based on a level probability distribution result corresponding to the tensor output by the trained probability generator. The invention can effectively reduce the training complexity and the network parameter number under the condition of ensuring the authenticity of the system, and achieves the aim of optimizing the probability distribution of the probability shaping signal.

Inventors

TIAN QINGHUA
ZHOU SITONG
XIN XIANGJUN
CAI YIFAN
GAO RAN
TIAN FENG
Dong ze
PAN XIAOLONG
ZHANG QI
WANG FU

Assignees

北京邮电大学

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (10)

1. A method for end-to-end optimization of a probability shaping signal, the method comprising: In each iteration period, receiving a disturbed signal which is transmitted and output by a transmitting end through a fiber channel, wherein the disturbed signal is obtained based on a signal obtained by carrying out electro-optical conversion on the transmitted signal, the transmitted signal is generated by carrying out constellation symbol mapping in advance on the basis of a sampled result and a standard constellation diagram, the sampled result is obtained by carrying out discretization sampling on a constellation point prior probability distribution characterization result, the constellation point prior probability distribution characterization result is obtained by inputting a preset tensor into a probability generator, and the level probability distribution result output by the probability generator is obtained, the probability generator comprises a nonlinear mapping cascade unit, a full connection layer and a softmax output layer, which are sequentially connected, and the nonlinear mapping cascade unit comprises a plurality of nonlinear mapping units which are sequentially connected; the demodulated signal obtained by carrying out photoelectric conversion and compensation recovery on the disturbed signal is input into a decoder, so that the decoder outputs a judgment result; And performing end-to-end joint training on the probability generator and the decoder through a joint transmitting end so that the transmitting end can obtain a target probability shaping signal based on a level probability distribution result corresponding to the tensor output by the trained probability generator.
2. The method according to claim 1, wherein the method further comprises: The method comprises the steps of obtaining a sending signal, inputting the sending signal and a demodulation signal into a mutual information estimator, enabling the mutual information estimator to output mutual information between the sending signal and the demodulation signal, and training the mutual information estimator, wherein the end-to-end joint training and the training of the mutual information estimator are alternately carried out, the mutual information estimator comprises a judging network, a variation distribution network and an operation module, the judging network and the variation distribution network comprise a plurality of first units and full-connection layers which are sequentially connected, each first unit comprises a full-connection layer and an activation function layer which are sequentially connected, and the operation module calculates the mutual information based on output results of the judging network and the variation distribution network.
3. The method of claim 2, wherein the parameters of the probability generator and the decoder are thawed in the case of end-to-end joint training of the probability generator and the decoder, the parameters of the mutual information estimator are frozen, gradient propagation during signal transmission is restored, and wherein the parameters of the mutual information estimator are thawed in the case of training of the mutual information estimator, the parameters of the probability generator and the decoder are frozen, and gradient propagation during signal transmission is cut off.
4. The method of claim 2, wherein the mutual information estimator is trained by optimizing a mutual information estimation loss as follows: Wherein, the Representing the estimated loss of mutual information, Representing the mutual information of the two-way communication, Representing the parameters of the trainable interpolation, And The regularization coefficient is represented as a function of the regularization coefficient, Representing the number of network layers of the mutual information estimator, Represent the first The number of parameters of the individual network layers, Represent the first First network layer of And a weight parameter.
5. The method of claim 4, wherein the probability generator and the decoder are trained end-to-end by optimizing joint loss as follows: Wherein, the Indicating the loss of the association, Representing the order of the QAM modulated signal, Representing the first of the constellation point prior probability distribution characterization results The prior probability distribution of the individual constellation points, Indicating the length of the transmitted signal, The actual posterior distribution is represented by a graph, The result of the decision is indicated, Representing the demodulated signal and, Representing the weight coefficient.
6. The method of claim 1, wherein the probability generator and the decoder are trained end-to-end by optimizing joint loss as follows: Wherein, the Indicating the loss of the association, Representing the order of the QAM modulated signal, Representing the first of the constellation point prior probability distribution characterization results The prior probability distribution of the individual constellation points, Indicating the length of the transmitted signal, The actual posterior distribution is represented by a graph, The result of the decision is indicated, Representing the demodulated signal.
7. A method for end-to-end optimization of a probability shaping signal, the method comprising: in each iteration period, inputting a preset tensor into a probability generator, and enabling the probability generator to output a level probability distribution result to obtain a constellation point prior probability distribution characterization result based on the level probability distribution result, wherein the probability generator comprises a nonlinear mapping cascade unit, a full connection layer and a softmax output layer which are sequentially connected, and the nonlinear mapping cascade unit comprises a plurality of nonlinear mapping units which are sequentially connected; Discretizing and sampling the characteristic result of the prior probability distribution of the constellation points, and performing constellation symbol mapping based on the sampled result and a standard constellation diagram to generate a transmission signal; And carrying out optical fiber channel transmission after carrying out electro-optical conversion on the transmitting signal so that a receiving end can receive a disturbed signal output by transmission, carrying out photoelectric conversion on the disturbed signal, then carrying out compensation and recovery on the disturbed signal, inputting the demodulated signal into a decoder, outputting a judgment result by the decoder, and carrying out end-to-end joint training on the probability generator and the decoder by a joint transmitting end so that the transmitting end can obtain a target probability shaping signal based on a level probability distribution result corresponding to the tensor output by the trained probability generator.
8. The method of claim 7, wherein after generating the transmit signal, the method further comprises processing the transmit signal to obtain a processed signal for optical fiber channel transmission after electro-optical conversion of the processed signal, the processing comprising up-sampling, pulse shaping, and random phase rotation in sequence, and the compensating recovering comprising random phase rotation, dispersion compensation, nonlinear equalization, matched filtering, down-sampling, equalization, and phase recovery in sequence.
9. A probability shaping signal end-to-end optimizing device comprising a processor, a memory and computer instructions stored on the memory, wherein the processor is adapted to execute the computer instructions, which when executed, implement the steps of the method according to any one of claims 1 to 8.
10. The end-to-end optimizing system of the probability shaping signal is characterized by comprising a transmitter, a fiber channel model and a receiver; The transmitter is used for inputting a preset tensor into a probability generator in each iteration period, enabling the probability generator to output a level probability distribution result so as to obtain a constellation point prior probability distribution representation result based on the level probability distribution result, performing discretization sampling on the constellation point prior probability distribution representation result, performing constellation symbol mapping based on the sampled result and a standard constellation diagram to generate a transmission signal, and performing electro-optical conversion on the transmission signal, wherein the probability generator comprises a nonlinear mapping cascade unit, a full connection layer and a softmax output layer which are sequentially connected, and the nonlinear mapping cascade unit comprises a plurality of nonlinear mapping units which are sequentially connected; the fiber channel model is used for carrying out fiber channel transmission on the converted signals in each iteration period and outputting disturbed signals; The receiver is used for receiving the disturbed signal in each iteration period, carrying out photoelectric conversion on the disturbed signal, and then compensating and recovering the disturbed signal to obtain a demodulation signal, and inputting the demodulation signal into the decoder to enable the decoder to output a judgment result; and the transmitter and the receiver are combined, and the probability generator and the decoder are subjected to end-to-end combined training, so that the transmitter can obtain a target probability shaping signal based on the level probability distribution result corresponding to the tensor output by the trained probability generator.

Description

End-to-end optimization method, device and system for probability shaping signal Technical Field The present invention relates to the field of optical fiber communications technologies, and in particular, to an end-to-end optimization method, apparatus, and system for a probability shaping signal. Background In the field of optical communication, higher transmission capacity, higher energy distribution efficiency and optimization of transmission performance of optical fiber communication are the targets pursued at present. QAM modulation is widely used in optical fiber communication systems as an effective means for improving the spectral efficiency of the system. The research shows that a 1.53dB gap exists between the uniformly distributed input high-order QAM signal and the Shannon limit, and the probability shaping technology can enable the probability distribution of constellation points to be more matched with a transmission channel by adjusting the probability distribution of the constellation points, so that the gap between the constellation points and the Shannon limit is reduced. The probability shaping technology and the QAM modulation technology are combined, so that the transmission capacity of an optical communication system can be effectively improved, and coherent digital signal processing can be realized through robust square QAM. For square QAM modulated signals, how to find the optimal probability distribution under the current transmission conditions is a research hotspot. The end-to-end optimization technology based on deep learning is a method which is paid attention to in artificial intelligence, an automation system and complex engineering tasks, a unified trainable model is built from original input to final output, a module of manual design or a staged processing flow is not relied on, and the whole model is subjected to joint optimization by using a gradient descent method. In recent years, end-to-end optimization technology based on deep learning has been widely applied to constellation shaping optimization, and geometric distribution and probability distribution are optimized by self-encoder architecture with mutual information or generalized normalized mutual information as targets. However, conventional end-to-end optimization techniques based on deep learning typically employ an output scheme of full constellation point mapping when probability shaping is performed for higher order QAM signals. Under this architecture, the number of neurons of the output layer is proportional to the constellation order, resulting in a linear increase in the number of network parameters with increasing modulation order. If a multi-layer complex network structure is adopted, the parameter scale can even be exponentially increased, and the calculation cost and the convergence difficulty of training are greatly increased. In addition, the existing end-to-end framework is often based on a simplified channel model, and the influence of a key DSP module in an actual coherent receiver is ignored, so that the model is easy to trap into a local optimal trap when nonlinear damage is processed. Therefore, how to construct an end-to-end optimization architecture with low computational complexity on the premise of ensuring the authenticity of the system is a key problem to be solved in the current optical communication field. Disclosure of Invention In view of this, embodiments of the present invention provide a method, apparatus, and system for end-to-end optimization of a probability-shaped signal to obviate or mitigate one or more disadvantages in the prior art. One aspect of the present invention provides a method for end-to-end optimization of a probability shaped signal, the method comprising the steps of: In each iteration period, receiving a disturbed signal which is transmitted and output by a transmitting end through a fiber channel, wherein the disturbed signal is obtained based on a signal obtained by carrying out electro-optical conversion on the transmitted signal, the transmitted signal is generated by carrying out constellation symbol mapping in advance on the basis of a sampled result and a standard constellation diagram, the sampled result is obtained by carrying out discretization sampling on a constellation point prior probability distribution characterization result, the constellation point prior probability distribution characterization result is obtained by inputting a preset tensor into a probability generator, and the level probability distribution result output by the probability generator is obtained, the probability generator comprises a nonlinear mapping cascade unit, a full connection layer and a softmax output layer, which are sequentially connected, and the nonlinear mapping cascade unit comprises a plurality of nonlinear mapping units which are sequentially connected; the demodulated signal obtained by carrying out photoelectric conversion and compensation recovery on the di