Neural network equalization method and system based on principal component time sequence

Abstract

The invention belongs to the technical field of optical fiber communication, and particularly relates to a neural network equalization method and system based on a principal component time sequence, and the method comprises the steps: carrying out the feature extraction of a signal principal component based on a perturbation theory; performing channel equalization based on a time sequence neural network; for a wavelength division multiplexing multi-path transmission coherent optical communication system, transmission signal data is obtained through a transmission system, preprocessing disturbance is performed on signals, then channel equalization is realized according to extracted disturbance signal principal component characteristic values and neural network estimation, a reliable and stable communication system can be provided, and the communication efficiency is improved. According to the invention, first-order perturbation is carried out on transmitted data and a signal principal component value is extracted, so that key prior information and nonlinear damage characteristics can be provided for subsequent channel equalization; and the extracted feature information is trained and predicted based on the time sequence neural network, so that the dynamic change of the signal can be captured in real time and the channel characteristics can be quickly adapted, and the nonlinear damage can be effectively compensated.

Description

Technical Field

[0001] The present invention relates to the field of optical communication, and particularly to a principal component time-series neural network equalization method and system. Background Art

[0002] With the rapid development of Internet and big data services, the research on fiber optic communication technologies with high speed, large capacity, and long distance is the current trend. Due to the existence of non-linear impairments in the channel, the performance of long-distance transmission systems is greatly affected. Based on the above background, channel equalization of the transmission channel is the key to improving system performance. In recent years, the development of multi-core high-performance processors and the continuous improvement of dedicated signal processing chips such as have provided powerful hardware support for the complex operations of channel equalization algorithms. This enables some channel equalization algorithms with large computational requirements to be more widely applied in actual communication systems to recover signals affected by noise in complex transmission environments.

[0003] An efficient and low-computational-complexity channel equalization method is crucial for transmission systems far from complex transmission conditions. Many traditional channel equalization methods are widely known. Such as the maximum likelihood sequence estimation algorithm (MLSE). The MLSE algorithm is an equalization method based on the maximum likelihood criterion, which considers all possible transmitted sequences, calculates the degree of matching between the received signal and each possible transmitted sequence after passing through the channel, and selects the sequence with the highest degree of matching as the estimated transmitted sequence. It has a good equalization effect and significantly improves system performance. However, the computational complexity of this method is relatively high, and usually, the Viterbi algorithm needs to be used to reduce the complexity. In addition, a linear equalizer estimates the signal through a linear filter, which consists of a plurality of tapped delay lines with weighted coefficients in a transversal filter structure. These weighted coefficients are adjusted to filter the received signal to compensate for the distortion impairments of the channel. Some of these traditional algorithms require prior knowledge of the transmission link parameters, while others are difficult to meet the requirements of real-time processing due to high computational complexity. Currently, there is still a lack of a channel equalization method with low complexity and effective compensation for non-linear impairments. Summary of the Invention

[0004] The present invention provides a principal component time-series neural network equalization method and system. The present invention is a channel equalization method with low spatial complexity and good equalization effect, which can effectively equalize the channel in a coherent optical fiber communication system.

[0005] Performing a first-order micro-perturbation on the transmitted data and extracting the principal component values of the signal can provide key prior information and non-linear impairment characteristics for subsequent channel equalization; based on the time-series neural network to train and predict the extracted feature information, it can capture the dynamic changes of the signal in real time and quickly adapt to the channel characteristics, thereby effectively compensating for non-linear impairments, and has great potential and application scenarios in the field of optical fiber communication.

[0006] The above technical object of the present invention is achieved by the following technical solutions:

[0007] A principal component time series neural network equalization method and system, comprising:

[0008] Perform perturbation analysis and principal component extraction according to the transmitted data. The specific steps are as follows:

[0009] (1) Collect the coherent transmitted signal. The time-distance propagation equation of the optical pulse in the optical fiber can be described by the nonlinear Schrödinger equation:

[0010] Where, α represents the linear attenuation, β2 represents the group velocity dispersion, γ represents the nonlinear effect, and U x / y (z,t) represents the complex signal envelope of x and y polarizations at the delay time frame t and distance z in the optical fiber;

[0011] (2) Through the first-order perturbation theory, derive the nonlinear perturbation term, expressed as:

[0012] Where, ΔU x and ΔU y are the nonlinear perturbation terms in the x and y polarizations respectively, P0 represents the peak power of the transmitted signal, C m,n represents the perturbation constant, which is determined by the actual parameters of the optical fiber, and T x and T y represent the perturbation triple;

[0013] (3) Extract the real part matrix of the signal, which can be expressed as S = Re(T p×q );

[0014] Where, T p×q represents the input p samples, the q-dimensional feature complex matrix, and calculate the covariance matrix

[0015] (4) Perform eigen-decomposition on the extracted features: Gh j = λ j h j , form the eigenvectors, which can be expressed as Re(T p×r ) = Re(T p×q )·H q×r Im(T p×r ) = Im(T p×q )·H q×r , and obtain the principal component feature matrix T p×q .

[0016] Preferably, the extracted signal is subjected to channel equalization using a temporal neural network, and the specific steps are as follows:

[0017] (1) Construct a reset gate that controls the previous time step hidden state for the current candidate hidden state: r t = σ(W r x t + U r h t-1 + b r );

[0018] Where h t-1 represents the previous time step hidden state, W r and U r represent weight matrices, b r is a bias vector, and σ represents an activation function; construct an update gate that controls the previous time step hidden state and the current candidate hidden state: z t = σ(W z x t + U z h t-1 + b z ) The current hidden state is a weighted sum of the previous hidden state and the candidate hidden state, and the weights are controlled by the update gate:

[0019] (2) Construct a gated recurrent unit: Where represents the hidden state of the forward gated recurrent unit processing data from the start to the end of the sequence, represents the hidden state of the reverse gated recurrent unit processing data from the end to the start of the sequence;

[0020] (3) Compare the equalization signal of the temporal neural network with the original signal, and calculate the difference between the equalization signal and the original signal:

[0021] Where n represents the number of samples, y represents the original signal, X represents the signal equalized by the neural network, and β represents the parameters of the temporal neural network.

[0022] In summary, the present invention mainly has the following beneficial effects:

[0023] The present invention is directed to a wavelength division multiplexing multi-channel coherent optical communication system. By obtaining transmission signal data through the transmission system, preprocessing and perturbing the signal, and then based on the extraction of the main component eigenvalues of the perturbed signal and neural network estimation, channel equalization is achieved, capable of providing a reliable and stable communication system, and having great potential and application prospects in the field of optical fiber communication. Brief Description of the Drawings

[0024] Figure 1 This is the block diagram of the principal component time series neural network transmission system based on the coherent optical communication system in the present invention.

[0025] Figure 2 This is the schematic diagram of the algorithm for extracting the principal component features of the signal based on perturbation analysis in the present invention;

[0026] Figure 3 This is the flowchart of the steps of the channel equalization algorithm using the time series neural network;

[0027] Figure 4 This is the comparison diagram of the complexities of different channel equalization algorithms in Embodiment 1;

[0028] Figure 5 This is the comparison diagram of the effects of the 64QAM signal after different channel equalizations in Embodiment 1. Detailed implementation manners

[0029] In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, rather than all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] The following embodiments are used to illustrate the present invention, but cannot be used to limit the protection scope of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions. Any simple improvement to the method of the present invention under the premise of the concept of the present invention shall fall within the protection scope required by the present invention.

[0031] For the convenience of those skilled in the art to accurately understand the technical principle and implementation path of the present invention, the coherent optical communication system shown in Figure 1 is taken as a specific embodiment to systematically elaborate on the working mechanism of the present invention.

[0032] The following combines Figures 2 - 5 , and takes the 64QAM signal as an example to illustrate the specific implementation manners of the present invention.

[0033] Embodiment 1

[0034] To meet the growing network transmission requirements, the development of high-speed, large-capacity, and long-distance optical fiber communication technology has become inevitable. However, the nonlinear damage effects in optical fibers significantly limit the system performance and transmission distance. Therefore, channel equalization has become a key technology for improving the long-distance transmission quality.

[0035] In the present invention, a channel equalization method based on a principal component time series neural network for wavelength division multiplexing coherent optical fiber communication is designed. The overall DSP system includes: low-pass filtering, Gram-Schmidt orthogonalization, dispersion compensation, constant modulus algorithm equalization, frequency offset estimation, carrier phase recovery, and a specially designed nonlinear compensation module. The core innovation lies in extracting the nonlinear characteristics of the signal using the first-order perturbation theory and real-time equalizing the channel nonlinear damage through a GRU network to achieve efficient signal recognition and channel equalization with low complexity.

[0036] The specific algorithm process is as follows:

[0037] 1. Collect the DP-64QAM signal transmitted through wavelength division multiplexing polarization coherence. Based on the nonlinear Schrödinger equation, describe the evolution relationship of the optical field in time and propagation distance through the first-order perturbation theory:

[0038] Derive the nonlinear perturbation term through the first-order perturbation theory: At the receiving end, decompose the extracted signal features to form a feature vector, which can be expressed as Re(T p×r ) = Re(T p×q )·H q×r Im(T p×r ) = Im(T p×q )·H q×r ;

[0039] Obtain the feature matrix T p×q containing the main component features of the signal. To reduce the computational complexity, limit the values of m and n:

[0040] 2. Package the extracted symbols. The i-th symbol r i is packaged as x i = [r i-k , …, r i , … r i+k . Take the packaged vector as the input sequence of the time series neural network. First, set the recursive time of the GRU model layer based on the time step to 2k + 1. The hidden states and between the gated units contain the symbol information flow between the recursive times; further input the h t output by the GUR layer to the linear layer. The number of nodes in the linear layer is equal to the number of samples n of the QAM modulation format. The output of the linear layer is the equalized signal after being equalized by the time series neural network. Compare the 64-QAM equalized signal output by the time series neural network with the original signal

[0041] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A principal component time series neural network based equalization method and system, characterized in that Including: (1) Generating acquisition data after compensation through a coherent optical communication system; (2) Perturbing the acquisition data and extracting data features; (3) Performing channel equalization based on a temporal neural network.

2. The transmission data compensation generation according to claim 1, wherein The signal compensation template includes: (1) Low-pass filtering module: This module is connected to the coherent receiver, mainly filtering out high-frequency noise and interference components, retaining the low-frequency part of the signal, and improving the quality and stability of the signal; (2) Gram-Schmidt orthogonalization module: This module is connected to the low-pass filtering module, mainly performing orthogonalization processing on the received signal to eliminate the correlation between polarization states that may occur during signal transmission; (3) Dispersion compensation module: This module is connected to the Gram-Schmidt orthogonalization module, mainly compensating for signal distortion caused by dispersion effects during fiber optic transmission and restoring the original shape of the signal; (4) Constant modulus algorithm equalization module: This module is connected to the dispersion compensation module, mainly keeping the amplitude of the signal at a constant level by adaptively adjusting the weights of the equalizer; (5) Frequency offset estimation module: This module is connected to the constant modulus algorithm equalization module, mainly estimating and compensating for frequency offset during signal transmission; (6) Carrier phase recovery module: This module is connected to the frequency offset estimation module, mainly for recovering the carrier phase of the signal and eliminating phase errors during transmission; (7) Nonlinear equalization module: This module is connected to the carrier phase recovery module, mainly for compensating for signal distortion caused by nonlinear effects during fiber optic transmission.

3. Perturb the data and extract data features according to claim 1, characterized in that The specific steps are as follows: (1) Establish the nonlinear Schrödinger equation: (2) Based on the perturbation theory, the perturbation terms of the nonlinear distortion for x- and y-polarizations are as follows: (3) Calculate the covariance matrix of the real part data: (4) Solve the characteristic equation to obtain the characteristic matrix: Re(T p×r ) = Re(T p×q )·H q×r , Im(T p×r ) = Im(T p×q )·H q×r .

4. Channel equalization is performed using the temporal neural network according to claim 1, characterized in that The specific steps are as follows: (1)Construct a gated recurrent unit composed of a reset gate and an update gate: r t = σ(W r x t + U r h t-1 + b r ) z t = σ(W z x t + U z h t-1 + b z ); (2) Construct a gated recurrent unit: (3) Compare the obtained equalized signal with the original transmitted signal:

Images