Balancing method for probabilistic shaping signal

Abstract

The invention belongs to the technical field of coherent light transmission, and particularly relates to an equalization method for probabilistic shaping signals. The method comprises the following steps: grouping to-be-equalized signals according to an amplitude ring; assuming that the number of constellation points of an original QAM constellation is M, constellation points in the M-QAM constellation share a kinds of amplitudes which are r1, r2,..., ra from small to large, and n sub-equalizers are arranged in total, all signals are divided into n classes corresponding to the amplitudes of the signals and the sub-equalizers used during equalization; each group is processed by an equalizer, namely a sub-equalizer; different sub-equalizer models are used for carrying out equalizer training on data of different groups, so that the non-uniform degree of probability distribution of training data of a single sub-equalizer is reduced, and signal equalization with better performance is realized. According to the invention, on the premise of almost not increasing the complexity overhead, the PS signal equalization effect is improved, and the performance gain is realized. In addition, by adopting the design of a plurality of sub-equalizers, the parallelism of signal equalization can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of coherent optical transmission, and in particular relates to an equalization method for probability shaping signals. Background Art

[0002] With the increase in bandwidth demand, expanding the capacity of optical transmission has become a long-term demand. Probabilistic shaping (PS) of the transmitted optical signal can make the signal power distribution more adaptable to the channel conditions under the existing hardware conditions, improve the tolerance of the optical signal-to-noise ratio, and thus achieve capacity expansion. At the same time, it can flexibly adjust the signal source entropy, and can provide more flexible rate adjustment for transmission than the traditional standard quadrature amplitude modulation (QAM). The existing probability shaping technology provides the optimal power distribution under Gaussian noise channels and can obtain shaping gain. However, there are certain problems for optical fiber transmission under bandwidth-limited conditions. In the optical fiber transmission channel, the damage to the signal is not only Gaussian noise, but also complex memory linear / nonlinear noise, including the filtering effect caused by band limitation, nonlinear noise introduced by optoelectronic devices and optical fibers, and optical fiber dispersion; therefore, there are a series of digital signal processing algorithms at the receiving end to compensate for these damages, and linear and nonlinear equalization is a key link. This type of equalization algorithm mainly achieves signal equalization by constructing linear or nonlinear filters and then training the filter weights based on data-driven training. The PS signal has an uneven distribution probability of signal constellation points at different amplitudes in the inner and outer circles, which will affect the performance loss when training the filter, so that the expected shaping gain cannot be obtained under the same signal-to-noise ratio. Summary of the invention

[0003] The object of the present invention is to provide an equalization method for probability shaping signals with good signal equalization performance, so as to improve the effect of PS signal equalization and achieve performance gain.

[0004] The equalization method for probability shaping signal provided by the present invention firstly groups the equalized signal according to the amplitude ring. Specifically, the grouping is based on the amplitude of the signal. Assuming that the number of constellation points of the original QAM constellation is M, the constellation points in the M-QAM constellation have a total of a amplitudes, which are r from small to large. 1 ,r 2 ,…,r a , there are n sub-equalizers in total, so all signals need to be divided into n categories, corresponding to their amplitudes and the sub-equalizers used for equalization.

[0005] For example, Figure 1As shown in the figure, the constellation points of the probability shaping (PS) 64-QAM constellation diagram are 9, with a total of 9 amplitudes, that is, 9 amplitude rings. The basis for grouping is that probability shaping will make the distribution probabilities of signals of different amplitudes different. For single-point signals, the smaller the amplitude, the greater the probability; therefore, the middle amplitude ring is set as the grouping boundary, and all signals are divided into groups with different amplitudes to reduce the uneven distribution of the signal data set, thereby reducing the degree to which the equalizer training process is affected by uneven data. Therefore, an equalizer is required for each grouping, that is, a sub-equalizer. Different sub-equalizer models are used to train the equalizer for data of different groups to reduce the unevenness of the probability distribution of the training data of a single sub-equalizer and achieve better signal equalization; therefore, the specific steps of the equalization method for probability shaping signals are as follows:

[0006] (1) Initialization of sub-equalizer module;

[0007] There are n sub-equalizer modules for equalizing signals to be equalized with different amplitudes. The models of the sub-equalizers include but are not limited to linear filters (M. Kikuta, S. Beppu, T. Nagai, Y. Saito, T. Tsuritani and K. Igarashi, "Real-Time DSP for Adaptive FIR Filters Based on LMS Algorithm," 2019 24th OptoElectronics and Communications Conference (OECC) and 2019 International Conference on Photonics in Switching and Computing (PSC), Fukuoka, Japan, 2019, pp. 1-3, doi: 10.23919 / PS.2019.8817646.), Volterra filters (M. Kong et al., "800-Gb / s / carrier WDM Coherent Transmission Over 2000km Based on Truncated PS-64QAM Utilizing MIMO Volterra Equalizer," in Journal of Lightwave Technology, vol.40, no.9, pp.2830-2839,1May1,2022,doi:10.1109 / JLT.2022.3148336.), Neural Network T.Kamiyama, H.Kobayashi and K.Iwashita, "NeuralNetwork Nonlinear Equalizer in Long-Distance Coherent Optical TransmissionSystems," in IEEE Photonics Technology Letters, vol.33, no.9, pp.421-424, 1May1, 2021, doi:10.1109 / LPT.2021.3067341.) etc.

[0008] (2) Equalizer training; each equalizer includes a sliding window module, an amplitude decision module, and a coupling output module;

[0009] The sliding window module is used to generate an input signal window for each amplitude grouping; the length of the sliding window depends on the number of taps of the equalizer, and the number of taps needs to be determined by debugging such as specific system conditions and baud rate;

[0010] The amplitude decision module determines the sub-equalizer to which the input signal block (i.e., the signal from the sliding window) is sent according to its amplitude; specifically, since there is a decision boundary after grouping, the signal is determined to which group it belongs according to the amplitude and sent to the corresponding sub-equalizer. During the training process, the signal at the overlapping loop is sent to two adjacent sub-equalizers;

[0011] The coupled output module is used to feed back the output signal to each corresponding sub-equalizer module during the training process to calculate the loss value; specifically, the loss function used to train the sub-equalizer calculates the loss value according to the equalizer output signal and the reference signal (the original signal after synchronization). For example, the mean square error function completes the training of each sub-equalizer.

[0012] (3) Signal transmission; including:

[0013] an amplitude decision module shared with the amplitude decision module in (2), which, during actual transmission, uses the overlapping ring radius between the sub-equalizers as the decision boundary to perform amplitude decision on the signal to be equalized and determine the sub-equalizer to which it is sent;

[0014] A coupling output module shared with the coupling output module in (2) is used to summarize the outputs of the currently operating sub-equalizers and output the output signals of the total equalizer in sequence.

[0015] In addition, there can also be a coherent optical communication system for actual transmission, specifically, a coherent optical communication system for transmitting PS-QAM signals. Assuming that the number of constellation points of the QAM constellation is 9 (and the same applies if the number of constellation points is M), the following system can be used (see Figure 2Figure 1 shows the transmitter module, which consists of a commercial DAC with 20 GHz analog bandwidth and 84 GSa / s sampling rate, a 100 kHz linewidth external cavity laser (ECL), a 32 GHz bandwidth I / Q modulator, two 65 GHz modulator drivers, and a polarization multiplexer (Pol.MUX). For optical signal modulation, the 1553.126 nm light source from the ECL is fed into the I / Q modulator driven by the pre-processed signal. After optical modulation, polarization multiplexing is achieved using the Pol.MUX. On the transmitter side, the 8-channel 50GHz-WDM PM signal is divided into odd channel groups (channels 1, 3, 5, and 7) and even channel groups (channels 2, 4, 6, and 8), combined by polarization-maintaining optical couplers (PM-OCs) and launched into a cyclic ULAF loop with a reverse-pumped Raman amplifier, which consists of several sections of 400km TeraWave+ fiber with an average effective area of ​​125μm2, an attenuation coefficient of 0.182dB / km, and a dispersion coefficient of 20ps / (nm·km) at 1550nm. In the cyclic loop, a reverse-pumped Raman amplifier with a 20dB switching gain is used to compensate for the signal loss of each span. The average power of the Raman pump is about 950mW. In addition, we use an attenuator (ATT) to control the optical power and use WSS to flatten the gain slope of the bandpass filter. On the receiving end, a tunable optical filter (TOF) with a 3dB bandwidth of 0.9nm is used to select the desired sub-channel from the 8-channel WDM signal. In our experiment, the 4th subchannel is selected by TOF. Then, two polarization beam splitters (PBS) and two 90° optical hybrids are used to implement polarization and phase diversity operations on the selected WDM signal and 100kHz linewidth optical local oscillator (LO). After that, the output of the hybrid is fed to four balanced photodetectors (PD), each with a bandwidth of 70GHz. Finally, digitization and sampling are achieved using a digital oscilloscope (OSC) with a bandwidth of 65GHz and a sampling rate of 160GSa / s, followed by offline DSP. The offline DSP includes Butterworth digital low-pass filter, resampling, chromatic dispersion (CD) compensation, clock recovery, T / 2CMA, frequency offset estimation (FOE), carrier phase estimation (CPE) based on blind phase search (BPS), and the equalization method for probability-shaped signals proposed in this patent (using a neural network model as a sub-equalizer).

[0016] Assume that the number of constellation points of the original QAM constellation is M, and the constellation points in the M-QAM constellation have a total of a amplitudes, from small to large, respectively r 1 ,r 2 ,…,r a , there are n sub-equalizers in total, so all signals need to be divided into n categories, corresponding to their amplitudes and the sub-equalizers used for equalization.

[0017] By setting the coverage of the sub-equalizer with overlapping loops, the functionality of the sub-equalizer can be expanded, and the signal to be equalized can be prevented from being sent to the wrong sub-equalizer under the influence of noise in actual transmission. During the training process, the amplitude loops of the signal points to be equalized in the training data of each sub-equalizer have different sources, corresponding to n types. The signals at the overlapping loop positions will be sent to two adjacent sub-equalizers for training at the same time. In actual transmission, the amplitude of the overlapping loop becomes the decision boundary of the amplitude decision module, and the sub-equalizer to be used is determined according to the amplitude of the signal to be equalized.

[0018] Compared with the existing equalization technology, the present invention realizes the improved equalization of probability shaping signals by grouping signals of different amplitude loops, reduces the unevenness of each sub-equalizer training data; decomposes the equalization task into the equalization of signals at different amplitude loops, increases the pertinence of equalizer training; the signal of each sliding window still only passes through one (sub) equalizer module, and has a very low impact on complexity. At the same time, the present invention decomposes the equalization task through the amplitude judgment module, further improves the parallelism of the algorithm, and recovers the shaping gain lost by the probability shaping signal in the optical transmission system due to the poor convergence of the non-uniform training of the equalizer, providing a valuable reference for the equalization technology of future communication systems.

[0019] The present invention can improve the effect of PS signal equalization and achieve performance gain without increasing complexity overhead. In addition, the present invention uses a design of multiple sub-equalizers to help improve the parallelism of signal equalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of applying the equalization method for probability shaping signals (n=2) in the PS-64QAM format.

[0021] Figure 2 It is a block diagram of a coherent optical communication system for actual transmission of the present invention.

[0022] Figure 3 It is a transmission distance comparison diagram of the equalization method for probability shaping signals of the present invention when the sub-equalizer is a neural network model.

[0023] The numbers in the figure are: 1 is the input signal sequence, 2 is the sliding window module, 3 is the amplitude decision module, 4 is the inner circle equalizer, 5 is the outer circle equalizer, 6 is the overlapping ring, 7 is the coupling output module, 8 is the constellation diagram of the 64QAM modulation format and the annotation of each amplitude ring, and 9 is a set of probability statistics of signal points at each amplitude ring of a typical PS-64QAM signal. DETAILED DESCRIPTION

[0024] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0025] Figure 1 FIG. 1 is a schematic diagram showing the application of the probability shaping signal equalization method (n=2) in the PS-64QAM format, which includes:

[0026] In the training phase, the input signal (1) is intercepted by the sliding window module (2) as an equalizer input signal block through a sliding window. The amplitude decision module (3) determines the input sub-equalizer according to the amplitude ring where the signal to be equalized at the center of each signal block is located. In this example, the 64QAM constellation has a total of a = 9 amplitudes, corresponding to the 9 rings (ring1-ring9) shown in (8) and (9); the number of sub-equalizers is n = 2, so it can be simplified to say that one sub-equalizer is responsible for the equalization of the inner ring constellation points, and the other is used for the equalization of the outer ring constellation points. In this stage, the signal blocks to be equalized with the center signal amplitude located in the 1st to 5th ring will be sent to the inner ring equalizer (4), and the signal blocks to be equalized in the 5th to 9th ring will be sent to the outer ring equalizer (5). The sub-equalizer in this example is a neural network model, and its specific structure includes an input layer, a hidden layer, and an output layer. The input layer is used to input the signal block to be equalized obtained by the sliding window, the hidden layer and its nonlinear activation function between layers are used to fit the nonlinear effect of the channel, and the output layer is used to output the equalized signal. In the figure, x is the input signal block to be equalized, h is the hidden layer neuron, and y is the output equalized signal. The fifth ring is the overlapping ring (6), which will be sent to the two sub-equalizers for training at the same time during the training stage, so that both sub-equalizers have the ability to equalize the overlapping signals. The signal output by the sub-equalizer is fed back to the corresponding sub-equalizer through the coupling output module (7) in this stage for loss calculation and weight training.

[0027] In the actual transmission stage, the input signal (1) is intercepted by the sliding window module as an equalizer input signal block through the sliding window. The amplitude of the signal to be equalized at the center of each signal block is used to determine the input sub-equalizer through the amplitude decision module (3). The decision boundary is the amplitude r of the fifth ring. 5 , the central signal amplitude is less than r 5 The signal block is sent to the inner circle equalizer, and vice versa to the outer circle equalizer. The outputs of the two sub-equalizers are output through a coupling output module (7).

[0028] Those skilled in the art will appreciate that the above embodiments are specific examples for implementing the present invention, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present invention.

[0029] RW-NNE (Ring-Wise Neural Network Equalization) is an example of the equalization method for probability shaping signals of the present invention when the sub-equalizer is a neural network model. In the above transmission system, the PS-64QAM signal with a single-wave net rate of 420Gbps is transmitted. It can be seen that the bit error is significantly reduced, the performance of the equalizer limited by unevenly distributed data is compensated, more shaping gain is achieved, and the transmission distance is increased by 25% compared with ordinary neural network equalization. See Figure 3 shown.

Claims

1. An equalization method for a probability shaping signal, characterized in that: First, the equalized signals are grouped according to the amplitude ring; the basis for grouping is the amplitude of the signal. Assuming that the number of constellation points of the original QAM constellation is M, there are a types of amplitudes of constellation points in the M-QAM constellation, from small to large, they are r1, r2, …, r a , there are n sub-equalizers in total, then all signals are divided into n categories, corresponding to their amplitudes and the sub-equalizers used for equalization; each group is processed by an equalizer, namely a sub-equalizer; different sub-equalizer models are used to train the equalizer for data of different groups, so as to reduce the unevenness of the probability distribution of the training data of a single sub-equalizer and achieve better signal equalization; therefore, the specific steps of the equalization method for probability shaping signals are as follows: (1) Initialization of sub-equalizer module; There are n sub-equalizer modules for equalizing signals with different amplitudes, and the sub-equalizer model adopts a linear filter, a Volterra filter or a neural network; (2) Equalizer training; each equalizer includes a sliding window module, an amplitude decision module, and a coupling output module; The sliding window module is used to generate an input signal window for each amplitude grouping; the length of the sliding window depends on the number of taps of the equalizer, and the number of taps depends on the specific system conditions and baud rate debugging; The amplitude decision module determines the sub-equalizer to which the input signal block is sent according to the amplitude of the signal from the sliding window; specifically, since there is a decision boundary after grouping, the signal belongs to which group according to the amplitude, and is sent to the corresponding sub-equalizer. During the training process, the signal at the overlapping loop is sent to two adjacent sub-equalizers; The coupled output module is used to feed back the output signal to each corresponding sub-equalizer module during the training process to calculate the loss value; specifically, the loss function used to train the sub-equalizer calculates the loss value according to the equalizer output signal and the reference signal; and completes the training of each sub-equalizer; (3) Signal transmission; including use: An amplitude decision module shared with the amplitude decision module in step (2) is used to determine the amplitude of the signal to be equalized and determine the sub-equalizer to be sent to, using the overlapped ring radius between the sub-equalizers as the decision boundary during actual transmission; A coupling output module shared with the coupling output module in step (2) is used to summarize the outputs of the currently operating sub-equalizers and output the output signals of the total equalizer in sequence.

2. The equalization method for probability shaping signal according to claim 1, characterized in that: The signal transmission also includes using a coherent optical communication system for transmitting PS-QAM signals, assuming that the number of constellation points of the QAM constellation is 9, specifically including: The transmitter module consists of a commercial DAC with 20GHz analog bandwidth and 84GSa / s sampling rate, a 100kHz linewidth external cavity laser (ECL), a 32GHz bandwidth I / Q modulator, two 65GHz modulator drivers and a polarization multiplexer (Pol.MUX); for optical signal modulation, the 1553.126nm light source from the ECL is fed into the I / Q modulator driven by the pre-processed signal; after optical modulation, polarization multiplexing is achieved using Pol.MUX; on the transmitter side, the 8-channel 50GHz-WDM The PM signals are divided into odd channel groups: channels 1, 3, 5 and 7, and even channel groups: channels 2, 4, 6 and 8, combined by polarization-maintaining optical couplers (PM-OCs) and launched into a circulating ULAF loop with a reverse-pumped Raman amplifier, which consists of several 400km sections of TeraWave+ optical fiber with an average effective area of ​​125μm2, an attenuation coefficient of 0.182dB / km, and a dispersion coefficient of 20ps / (nm·km) at 1550nm; in the circulating loop, a reverse-pumped Raman amplifier with a 20dB switching gain is used to compensate for the signal loss of each span; the average power of the Raman pump is about 950mW; in addition, an attenuator (ATT) is used to control the optical power, and WSS is used to flatten the gain slope of the bandpass filter; At the receiving end, a tunable optical filter (TOF) with a 3dB bandwidth of 0.9nm is used to select the desired subchannel from the 8-channel WDM signal; the 4th subchannel is selected by the TOF; then, two polarization beam splitters (PBS) and two 90° optical hybrids are used to implement polarization and phase diversity operations on the selected WDM signal and a 100kHz linewidth optical local oscillator (LO); after that, the output of the hybrid is fed to four balanced photodetectors (PD), each with a bandwidth of 70GHz; finally, digitization and sampling are achieved using a digital oscilloscope (OSC) with a bandwidth of 65GHz and a sampling rate of 160GSa / s, followed by offline DSP; the offline DSP includes Butterworth digital low-pass filter, resampling, chromatic dispersion (CD) compensation, clock recovery, T / 2CMA, frequency offset estimation (FOE), carrier phase estimation (CPE) based on blind phase search (BPS), and equalization methods for probability-shaped signals.

Images