Modulation format recognition method for probabilistically shaped coherent optical communication system based on nonlinear signal power variation

Abstract

The present invention proposes a modulation format recognition scheme based on signal nonlinear power variation, and preprocesses the received probability shaping signal, including: uniformly normalizing the signal input to the modulation format recognition module, obtaining the signal amplitude information, obtaining the amplitude characteristics and spectrum characteristics by calculation, and inputting the extracted characteristic parameters as training data into a two-level fully connected deep neural network, thereby completing the recognition of the modulation format category of the PS-MQAM signal and the modulation format signal shaping entropy recognition. The present invention can simultaneously realize the modulation format recognition and shaping distribution recognition of the probability shaping signal containing multiple shaping parameter constellation probability distribution entropy intervals as narrow as 0.1 bit / symbol, and can also well achieve high recognition accuracy in the probability shaping coherent optical communication system under low optical signal-to-noise ratio (OSNR) conditions.

Description

Technical Field

[0001] The invention relates to the field of optical fiber communication, and in particular to a method for identifying modulation format categories and modulation format shaping entropy for a probability shaping coherent optical communication system. Background Art

[0002] The future optical fiber transmission network needs to meet not only the demand for ultra-high capacity, but also the complex and diverse business needs of users. This means that the next generation of optical networks needs to be able to monitor the communication link status and business service quality in real time, and dynamically adjust the modulation format and signal transmission rate of the transmitter according to the monitored data. In coherent transmission systems, the signal strength, phase and polarization characteristics vary greatly under different modulation formats, which makes the compensation and equalization algorithms under different modulation formats not exactly the same. Therefore, before executing these digital signal processing (DSP) algorithms, the modulation format information and shaping distribution information of the signal need to be obtained in advance. Therefore, in the optical fiber transmission network based on PS-MQAM, the flexible receiver must not only have the ability to identify the modulation format, but also the ability to identify the shaping distribution parameters.

[0003] In order to identify the above-mentioned debugging format information and shaping distribution information, researchers have proposed a modulation format recognition scheme based on Stokes space and a shaping distribution recognition scheme based on the maximum expectation algorithm of radius. The modulation format recognition and shaping distribution recognition schemes proposed above are relatively independent, and in the future flexible receivers, the ability to simultaneously identify these two types of information is required. Summary of the invention

[0004] The present invention provides a modulation format recognition method for a probability shaping coherent optical communication system, which can simultaneously complete the recognition of the signal modulation format category and the recognition of the modulation format shaping entropy.

[0005] In order to achieve the above objectives, the overall idea is: propose a scheme based on the nonlinear power change of the signal, perform uniform normalization processing on the signal input into the modulation format category recognition module, obtain the signal amplitude information and the fast Fourier transform spectrum information of the signal after 4, 8, and 16 powers, obtain the amplitude characteristics and spectrum characteristics related to the modulation format and shaping distribution by calculation, and input the extracted feature parameters as training data into a three-layer fully connected deep neural network with 256, 128, and 64 hidden layers.

[0006] Specifically, the present invention adopts a scheme for modulation identification and shaping distribution information identification of a probability-shaped coherent optical communication system based on nonlinear power variation of a signal, comprising the following steps:

[0007] The received PS-MQAM signal is calculated and normalized, where the PS-MQAM signal is a probability shaping quadrature amplitude modulation signal. The normalized signal is calculated in the following way:

[0008]

[0009] in, represents the preprocessed signal sequence, |·| represents the modulo operation, and max() represents the maximum value operation.

[0010] After calculating the nonlinear power conversion characteristics of the normalized signal, the PS-MQAM signal is processed to the nth power, and then Fourier transform is performed to calculate the frequency domain characteristics, through the following calculation method:

[0011]

[0012] Where M represents the n-th power processing of the normalized signal, and N represents the length of the Fourier transform.

[0013] Preferably, the value of M is 4, 8, and 16 respectively.

[0014] Preferably, the value of N is 2 16 .

[0015] The amplitude mean of the original signal Calculated by the following formula:

[0016]

[0017] Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points.

[0018] The amplitude variance σ of the original signal is calculated by the following formula:

[0019]

[0020] Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points.

[0021] The peak value P of the spectrum amplitude of the original signal after M power M Calculated by the following formula:

[0022] P M =max[Y M (k)].

[0023] Among them, k represents the sampling signal number, and max represents the maximum value operation.

[0024] The variance σ of the spectrum amplitude of the original signal after M-th power M Calculated by the following formula:

[0025]

[0026] The mean value of the spectrum amplitude after the original signal is raised to the Mth power E M Calculated by the following formula:

[0027]

[0028] The DNN model consists of an input layer, three hidden layers and an output layer, wherein the number of input layers is the same as the number of extracted amplitude features, the three hidden layers are composed of 256, 128 and 64 neurons respectively, the number of neurons contained in the output layer is the same as the number of modulation format types or shaping distribution types of the pre-identified signal, and the loss function uses the cross entropy loss function.

[0029] The amplitude mean of the original signal and amplitude variance σ and the spectrum amplitude peak P after the signal is raised to the Mth power M , variance σ M and mean E M Input into the first-level fully connected DNN model for modulation format recognition.

[0030] The amplitude mean of the original signal for each modulation format recognized by the first-level fully connected DNN model and amplitude variance σ and the spectrum amplitude peak P after the signal is raised to the Mth power M , variance σ M and mean E M Input into the second-level fully connected DNN model for shaping distribution recognition.

[0031] Compared with the prior art, the beneficial effects of the present invention are that a modulation format recognition scheme based on Stokes space and a shaping distribution recognition scheme based on a radius maximum expectation algorithm are relatively independent and cannot simultaneously complete the recognition of the modulation format and shaping distribution information. The present invention proposes a scheme based on the nonlinear power change of the signal according to the characteristics of the nonlinear power change of the probability shaping signal, obtains the signal amplitude information and the fast Fourier transform spectrum information of the signal after the 4th, 8th and 16th powers, obtains the amplitude characteristics and spectrum characteristics related to the modulation format category and the modulation format shaping entropy through calculation, and then inputs them into a two-level fully connected neural network for feature extraction, and finally completes the modulation format recognition of the probability shaping signal containing multiple shaping parameters at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of a digital signal processing (DSP) flow of modulation format and shaping distribution identification of the present invention;

[0033] Figure 2 It is a schematic diagram of the execution flow of the scheme for modulation format and shaping distribution identification of the present invention;

[0034] Figure 3 is a spectrum characteristic diagram of nonlinear power conversion of the modulation format of the present invention;

[0035] Figure 4 It is a schematic diagram of the correlation of the signal's nonlinear power variation characteristics;

[0036] Figure 5 is a schematic diagram of the two-stage DNN structure used in the present invention;

[0037] Figure 6 is a modulation format category recognition result diagram of a probability shaping signal;

[0038] Figure 7 It is a modulation format shaping entropy recognition result diagram of the probability shaping signal; DETAILED DESCRIPTION

[0039] In order to explain the embodiments and the specific implementation and advantages of the present invention in more detail, further description is given below in conjunction with the accompanying drawings:

[0040] A method for identifying modulation format and shaping distribution of a probability-shaped coherent optical communication system based on nonlinear power variation of a signal, the device used is as shown in the attached Figure 1 As shown, at the transmitting end, the bit stream is modulated by a Mach-Zehnder modulator and then enters the optical fiber. After long-distance transmission, it is received by a receiver, and the optical signal is converted into an electrical signal through photoelectric conversion. After dispersion compensation and polarization demultiplexing in the electrical domain, the modulation format of the signal and the shaping distribution information are identified using the present invention.

[0041] Execution process such as Figure 2 As shown, the received PS-MQAM signal is calculated and normalized, wherein the PS-MQAM signal is a probability shaping quadrature amplitude modulation signal, and the normalized signal is calculated in the following manner:

[0042]

[0043] in, represents the preprocessed signal sequence, |·| represents the modulo operation, and max() represents the maximum value operation.

[0044] After calculating the nonlinear power conversion characteristics of the normalized signal, the PS-MQAM signal is processed to the nth power, and then Fourier transform is performed to calculate the frequency domain characteristics, through the following calculation method:

[0045]

[0046] Where M represents the n-th power processing of the normalized signal, and N represents the length of the Fourier transform.

[0047] Preferably, the value of M is 4, 8, and 16 respectively.

[0048] Preferably, the value of N is 2 16 .

[0049] The amplitude mean of the original signal Calculated by the following formula:

[0050]

[0051] Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points.

[0052] The amplitude variance σ of the original signal is calculated by the following formula:

[0053]

[0054] Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points.

[0055] The peak value P of the spectrum amplitude of the original signal after M power M Calculated by the following formula:

[0056] P M =max[Y M (k)].

[0057] Among them, k represents the sampling signal number, and max represents the maximum value operation.

[0058] The variance σ of the spectrum amplitude of the original signal after M-th power M Calculated by the following formula:

[0059]

[0060] The mean value of the spectrum amplitude after the original signal is raised to the Mth power E M Calculated by the following formula:

[0061]

[0062] The DNN model is composed of an input layer, three hidden layers and an output layer, wherein the number of input layers is the same as the number of extracted amplitude features, the three hidden layers are composed of 256, 128 and 64 neurons respectively, and the output layer contains the number of neurons and the type of modulation format of the pre-identified signal (given by y 1 ,y 2 ,y 3 ,y 4represents the category of the first-level DNN output signal) or the number of shaping distribution types (respectively represented by P 1 , P 2 , P 3 , P 4 , P 5 , P 6 , P 7 and P 8 The loss function of the entire network training uses the cross-entropy loss function to process the output results so that the sum of the predicted values ​​of multiple categories is 1, and then the loss is calculated by cross-entropy. Generally, the smaller the cross-entropy value (Loss), the better the model prediction effect. The activation functions of the input layer and the three hidden layers are "Tanh" and "Relu" respectively, and the activation function of the output layer is "Softmax". The "Nadam" optimizer is used to speed up the convergence of the model.

[0063] The amplitude mean of the original signal and amplitude variance σ and the spectrum amplitude peak P after the signal is raised to the Mth power M , variance σ M and mean E M Input into the first-level fully connected DNN model for modulation format recognition.

[0064] The amplitude mean of the original signal for each modulation format recognized by the first-level fully connected DNN model and amplitude variance σ and the spectrum amplitude peak P after the signal is raised to the Mth power M , variance σ M and mean E M Input into the second-level fully connected DNN model for shaping distribution recognition.

[0065] Figure 3 The following is a schematic diagram of the amplitude distribution of PS-MQAM signals with different modulation formats and the spectrum distribution after Fourier transformation. Figure 3 (a) It can be seen that the amplitude distribution under different modulation formats is quite different. In order to obtain the amplitude distribution difference related to the modulation format, this scheme obtains the amplitude characteristics by calculating the amplitude mean and amplitude variance of the overall signal. Figure 3(a), (c), and (d) show the spectrum amplitude characteristics of the signal after the 4th, 8th, and 16th powers, respectively. The spectrum of PS-16QAM in all three states has peaks, while PS-64QAM has no obvious peaks in the three cases. As the power increases, the peaks of PS-32QAM and PS-128QAM appear from nothing to something, thus reflecting that the modulation format information is related to the signal peaks after different powers. In addition, it can be observed that there are differences in the amplitude of the spectrum changes of signals with different modulation formats, which can also be reflected by the amplitude mean and amplitude variance of the statistical spectrum.

[0066] Figure 4 The above-mentioned PS-16QAM, PS-32QAM, PS-64QAM and PS-128QAM are all characteristic related schematic diagrams. It can be seen from the figure that all the features are closely related to the modulation format and shaping distribution information of the signal, which illustrates the feasibility of the signal modulation format identification scheme using nonlinear power variation characteristics.

[0067] Figure 5 It is a schematic diagram of the DNN structure used in the present invention, where each level of the DNN model consists of an input layer, three hidden layers and an output layer, wherein the number of input layers is the same as the number of extracted amplitude features, the three hidden layers are composed of 256, 128 and 64 neurons respectively, the output layer contains the same number of neurons as the modulation format type or shaping distribution type of the pre-identified signal, the activation functions of the input layer and the three hidden layers are "Tanh" and "Relu" respectively, and the activation function of the output layer is "Softmax" as the classified output.

[0068] The feasibility of the modulation format recognition and shaping distribution recognition method of the probability shaping signal based on the nonlinear power variation characteristics of the signal is verified through simulation data, and the recognition accuracy of the method is verified in the transmission link. The simulation data contains 28GBaud PDM-PS-MQAM signals under 8 different shaping distributions, where the signal constellation entropy range is PS-16QAM[3.1,3.8], PS-32QAM[4.1,4.8], PS-64QAM[4.9,5.6], PS-128QAM[5.7,6.4], and the constellation entropy granularity is 0.1bit / symbol, respectively. 1 , P 2 , P 3 , P 4 , P 5 , P 6 , P 7 and P 8The OSNR (Optical Signal-to-Noise Ratio) of the simulation system ranges from 10dB to 32dB, 12dB to 34dB, 14dB to 36dB, and 16dB to 38dB, with a step size of 1dB. 200 sets of independent data are collected for each OSNR value, and each set of data contains 20,000 symbols. 70% of the data for each OSNR value is randomly selected to train the DNN network structure, and the remaining 30% of the data is used to test the recognition accuracy of the DNN model.

[0069] The grouped training set is imported into the DNN model, and the result information after the first learning is output. The initial learning rate (0.001) of each group of training sets is calculated based on the result information. The output results after the first learning are imported into the DNN model again for multiple iterations, and the loss functions corresponding to different groups of training sets are determined, so that the loss functions are linearly related, and the output results of different groups of training sets after multiple iterations are derived. The output result errors of different groups of training sets are judged, and the learning is stopped when the output result error is less than the preset threshold. The recognition accuracy of the DNN model is judged according to the test set, and the output result of the trained DNN model is obtained. The output result is compared with the test set for analysis to generate a deviation rate. If the deviation rate is less than the preset deviation rate threshold, it proves that the model prediction accuracy reaches the target value and the DNN model training is completed.

[0070] Figure 6 This is the modulation format result diagram identified by the first-level DNN network model. This method can realize the modulation format recognition of PS-MQAM signals with mixed multiple shaping parameters, and can achieve 100% modulation format recognition accuracy at the NGMI threshold where the signal OSNR value is lower than 0.935.

[0071] Figure 7 This is the result diagram of the shaping distribution information identified by the second-level DNN network model. This method can achieve 100% recognition at the NGMI threshold when the PS-MQAM signal constellation probability distribution entropy interval is 0.1 bit / symbol and the signal OSNR value is lower than 0.935.

[0072] These results show that the proposed method can work well under low optical signal-to-noise ratio (OSNR) conditions. The characteristic scheme based on the nonlinear power variation of the signal simultaneously completes the modulation format recognition and shaping distribution information recognition of the probability-shaped signal, thereby improving the performance of the probability-shaped coherent optical communication system.

Claims

1. A method for identifying modulation formats of a probability-shaped coherent optical communication system based on nonlinear power variation of a signal, characterized in that: The following steps are involved: The received PS-MQAM signal is normalized, where the PS-MQAM signal is a probability-shaped quadrature amplitude modulation signal. The normalized signal is calculated in the following way: Where y(n) represents the received PS-MQAM signal, represents the preprocessed signal sequence, represents the modulo operation, represents the maximum value operation, Calculate the nonlinear power spectrum characteristics of the normalized signal. The specific process includes: first, the normalized signal PS-MQAM signal Perform M-power processing Then Perform Fourier transform to obtain the corresponding nonlinear power spectrum characteristics. The calculation process is as follows: Where M represents the Mth power processing of the normalized signal, and N represents the length of the Fourier transform; Calculate the normalized signal The mean amplitude Calculated by the following formula: Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points; Calculate the normalized signal The amplitude variance σ is calculated by the following formula: Wherein, k represents the sampling signal number, and N represents the total number of sampling signal points; Calculate the nonlinear power spectrum Y M The peak value P of (k) M , calculated by the following formula: P M =max[Y M (k)], Where k represents the sampling signal number, Indicates the maximum value operation; Calculate the nonlinear power spectrum Y M The variance σ of (k) M , calculated by the following formula: Calculate the nonlinear power spectrum Y M The mean value E of (k) M , calculated by the following formula: The amplitude mean y and amplitude variance σ of the normalized signal and the spectrum amplitude peak value P after M power are M , variance σ M and mean E M Input into the first-level fully connected DNN model for modulation format recognition; For each modulation format recognized by the first-level fully connected DNN model, the amplitude mean y and amplitude variance σ of the normalized signal and the spectrum amplitude peak P after the signal is raised to the Mth power are calculated. M , variance σ M and mean E M Input into the second-level fully connected DNN model for shaping distribution recognition; The DNN model consists of an input layer, three hidden layers and an output layer, wherein the number of input layers is the same as the number of extracted amplitude features, the three hidden layers are composed of 256, 128 and 64 neurons respectively, the number of neurons in the output layer is the same as the number of modulation format types or shaping distribution types of the pre-identified signal, and the loss function is calculated using the cross entropy loss function (CrossEntropy Loss) by the following formula: Among them, p(x i ) represents the true value of the sample, q(x i ) represents the predicted value of the sample, and the amplitude feature refers to the amplitude mean and amplitude variance of the normalized signal as well as the spectrum amplitude peak, variance and mean.

Images