Optical fiber nonlinear equalization method and system based on bit-by-bit estimation neural network

Abstract

The application relates to the technical field of optical fiber communication, and particularly discloses an optical fiber nonlinear equalization method and system based on a bit-by-bit estimation neural network. The method comprises the following steps: acquiring M-QAM data of a coherent optical communication system sender, converting the M-QAM data into bits corresponding to a modulation format of the M-QAM data, and taking the bits as neural network labels; performing linear equalization on a signal transmitted through the coherent optical communication system to obtain data only retaining a nonlinear effect, and taking the data as neural network data; processing the data, constructing a feature sequence corresponding to the data, and constructing a training set and a test set based on the feature sequence and the labels; constructing a CNN-BiGRU neural network based on bit-by-bit estimation, training and optimizing the neural network by using the training set; and predicting corresponding sender data based on the test set by using the trained neural network. The method improves the optical fiber nonlinear equalization performance and reduces the complexity, and can be applied to a long-distance, high-order modulation format and a multi-channel system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical fiber communication technology, in particular to an optical fiber nonlinear equalization method and system based on bit-by-bit estimation neural network. BACKGROUND

[0002] In the process of developing coherent optical communication systems towards long distance and large capacity, high-speed optical signals will be affected by linear and nonlinear impairments from devices and optical fibers during transmission, resulting in degradation of system transmission performance. Linear compensation algorithms such as dispersion compensation, frequency offset estimation, and phase recovery algorithms are already relatively mature and can achieve good results. However, as the rate increases, the impact of nonlinear effects on signals becomes more and more significant, so optical fiber nonlinear equalization is of great significance for the further development of coherent optical communication systems.

[0003] As a widely used technology, neural networks have attracted increasing attention in the field of optical communication and network as the amount of data has exploded. Neural network-aided optimization of nonlinear equalization is an important research direction.

[0004] Neural networks can almost simulate any complex function when they have sufficient computing power and are trained on sufficient data sets, so they are particularly effective in solving complex nonlinear problems, making them a natural tool for optical fiber nonlinear equalization.

[0005] However, existing neural networks mostly use classification or regression methods to equalize nonlinearity. The classification method has too high a complexity when facing high-order modulation formats, and the regression method may be limited in modeling the details of the signal because regression is performed by minimizing the difference between the predicted output and the true value. High-order modulation formats, such as 64-QAM or 256-QAM, contain more levels of information in the amplitude and phase of the signal, which makes the nonlinear characteristics of the signal more pronounced, and thus requires neural network models with higher expressive power to capture these details. SUMMARY

[0006] To solve the above problems of existing neural network compensation methods, the present application provides an optical fiber nonlinear equalization method and system based on bit-by-bit estimation neural network, which uses bit-by-bit estimation instead of classification or regression to build neural networks, improving the performance of nonlinear equalization and reducing the complexity, and making the neural network more suitable for high-order modulation formats.

[0007] The technical solutions adopted by the present application are as follows:

[0008] An optical fiber nonlinear equalization method based on bit-by-bit estimation neural network, comprising the following steps:

[0009] S1, obtaining the M-QAM data of the coherent optical communication system, converting it into the corresponding bits of its modulation format as the neural network label;

[0010] S2, linearly equalizing the signal transmitted through the coherent optical communication system to obtain data only retaining the nonlinear effect as the neural network data;

[0011] S3, processing the data obtained by S1 and S2 to construct the feature sequence and label corresponding to the data, and constructing the training set and test set based on the feature sequence and label;

[0012] S4, constructing a CNN-BiGRU neural network based on bit-by-bit estimation, training and optimizing the neural network using the training set of S3;

[0013] S5, using the trained neural network of S4 to predict the corresponding transmitter data based on the test set of S3.

[0014] Preferably, S1 is specifically:

[0015] Obtaining the M-QAM signal of the coherent optical communication system, converting each symbol into corresponding binary bits according to its encoding method, each M-QAM signal corresponds to log2 M bits, and these bits are combined as Y i =[b1,b2,…,b log2(M) ], Yi is the neural network label.

[0016] Preferably, the linear equalization in S2 is specifically:

[0017] Processing the signal transmitted through the coherent optical communication system, including: IQ orthogonalization processing, dispersion compensation, frequency offset estimation, RRC low-pass filtering, downsampling, polarization demultiplexing, and phase estimation.

[0018] Preferably, S3 is specifically:

[0019] S3.1, dividing the x, y polarization state signals x i , y i obtained by S2 into real and imaginary parts and combining them into a data r i , considering the k data before and after, combining them into a feature sequence X i ;

[0020] S3.2, constructing the training set and test set using the feature sequence X i and the neural network label Y i obtained by S1.

[0021] Preferably, S3.1 is specifically:

[0022] For the complex-valued signal in the x polarization direction, it is decomposed into xi = [I x ,Q x ] for the complex-valued signal in the y polarization direction, decompose it into y i = [I y ,Q y ], and then combine them into data r i = [I x ,Q x ,I y ,Q y ];

[0023] Consider k data before and after r i and combine them to get the feature sequence X i = [r i-k ,…,r i-1 ,r i ,r i+1 ,…,r i+k ].

[0024] Preferably, S3.2 is specifically:

[0025] Use the feature sequence X i and the neural network label Y i to construct a training set {X i ,Y i};

[0026] And use new data to construct a test set according to S1-S3.

[0027] Preferably, S4 is specifically:

[0028] Build a CNN-BiGRU neural network based on bit-by-bit estimation, including: input layer, CNN layer, BiGRU layer, flat layer, fully connected layer, Sigmoid layer, output layer;

[0029] Based on the back propagation algorithm and the Adam optimizer, train the CNN-BiGRU neural network based on bit-by-bit estimation using the training set X i ,Y i}, and optimize each parameter of the neural network by minimizing the loss function.

[0030] Preferably, the loss function is a BCE Loss function, i.e. a binary cross-entropy loss function, which is specifically represented as:

[0031]

[0032] Where a i is a binary label 0 or 1, p(a i ) is the probability of output belonging to the label, and N represents the number of groups of prediction objects.

[0033] Preferably, S5 is specifically:

[0034] The predicted label corresponding to the current feature sequence is obtained by predicting the test set data through the neural network trained and optimized in S4, that is, the predicted bit corresponding to the M-QAM signal from the sending end.

[0035] A fiber nonlinear equalization system based on a bit-by-bit estimation neural network comprises the following modules:

[0036] A label acquisition module acquires M-QAM data from a coherent optical communication system, converts the M-QAM data into bits corresponding to the modulation format of the M-QAM data, and takes the bits as neural network labels.

[0037] A data acquisition module performs linear equalization on a signal transmitted through a coherent optical communication system to obtain data that only retains nonlinear effects, and takes the data as neural network data.

[0038] A sequence construction module processes the neural network data obtained by the data acquisition module to construct a feature sequence corresponding to the data.

[0039] A data set construction module constructs a training set and a test set based on the feature sequence of the sequence construction module and the neural network labels of the label acquisition module.

[0040] A model construction module constructs a CNN-BiGRU neural network based on bit-by-bit estimation.

[0041] A model training module trains and optimizes the neural network using the training set of the data set construction module.

[0042] A model prediction module predicts corresponding data from a sending end based on the test set of the data set construction module using the trained neural network of the model training module.

[0043] The present application has the following characteristics and advantages:

[0044] The present application discloses a fiber nonlinear equalization method and system based on a bit-by-bit estimation neural network, which acquires M-QAM data from a coherent optical communication system, converts the M-QAM data into bits corresponding to the modulation format of the M-QAM data, and takes the bits as neural network labels; performs linear equalization on a signal transmitted through a coherent optical communication system to obtain data that only retains nonlinear effects, and takes the data as neural network data; processes the data to construct a feature sequence corresponding to the data, and constructs a training set and a test set based on the feature sequence and the labels; constructs a CNN-BiGRU neural network based on bit-by-bit estimation, trains and optimizes the neural network using the training set; and predicts corresponding data from a sending end based on the test set using the trained neural network. The present application improves the performance of fiber nonlinear equalization, and is applicable to long-distance, high-order modulation format, and multi-channel systems. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flow chart of the optical fiber nonlinear equalization method based on bit-by-bit estimation neural network of the present application;

[0046] Figure 2 is a CNN-BiGRU neural network structure diagram based on bit-by-bit estimation and classification and regression methods proposed by the present application;

[0047] Figure 3 is a 90-channel WDM-PDM-16QAM experimental system structure diagram of the present application;

[0048] Figure 4 is a receiving end DSP flow chart of the experimental system of the present application;

[0049] Figure 5 is a Q factor performance comparison diagram of 32GBaud PDM-16QAM signals in a 90-channel WDM experimental system of the present application. DETAILED DESCRIPTION

[0050] The specific embodiments of the present application are described below to facilitate understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all inventions utilizing the concept of the present application are within the scope of protection.

[0051] As shown in Figure 1 , the present application provides an optical fiber nonlinear equalization method based on bit-by-bit estimation neural network, comprising the following steps:

[0052] Obtaining the M-QAM data of the transmitting end of the coherent optical communication system, converting it into the corresponding bits of its modulation format, and taking it as the neural network label;

[0053] Performing linear equalization on the signal transmitted through the coherent optical communication system to obtain data that only retains the nonlinear effect, as neural network data;

[0054] Processing the data, constructing the feature sequence corresponding to the data, and constructing the training set and test set based on the feature sequence and the label;

[0055] Constructing a CNN-BiGRU neural network based on bit-by-bit estimation, training and optimizing the neural network using the training set;

[0056] Using the trained neural network to predict the corresponding transmitting end data based on the test set.

[0057] Specifically: obtaining a coherent optical communication system transmitting end M-QAM signal, converting each symbol into corresponding binary bits according to its encoding mode, each M-QAM signal corresponds to log2(M) bits, and these bits are combined as Yi is the label required by the neural network.

[0058] The signal transmitted by the coherent optical communication system is processed, including IQ quadrature processing, dispersion compensation, frequency offset estimation, RRC low-pass filtering, downsampling, polarization demultiplexing, and phase estimation.

[0059] The x and y polarization state signals xi and yi are divided into real and imaginary parts and combined into one data ri. Considering the k data before and after, they are combined into a feature sequence Xi, and the training set and test set are constructed using the feature sequence Xi and the label Yi, including:

[0060] For the complex-valued signal in the x polarization direction, it is decomposed into xi=[Ix, Qx], and for the complex-valued signal in the y polarization direction, it is decomposed into yi=[Iy, Qy], and then they are combined into data ri=[Ix, Qx, Iy, Qy];

[0061] Considering the k data before and after ri and combining them to obtain the feature sequence Xi=[r i-k ,…r i-1 ,r i ,r i+1 ,…,r i+k ], the training set {X i ,Y i} is constructed using the feature sequence Xi and the label Yi, and the test set is constructed using new data according to the above process.

[0062] A CNN-BiGRU neural network based on bit-by-bit estimation is constructed as shown in Figure 2 , including an input layer, a CNN layer, a BiGRU layer, a flat layer, a fully connected layer, a Sigmoid layer, and an output layer.

[0063] Based on the back propagation algorithm and the Adam optimizer, the CNN-BiGRU neural network based on bit-by-bit estimation is trained using the training set X i ,Y i}, and the parameters of the neural network are optimized.

[0064] The test set data is predicted by the trained and optimized neural network, and the predicted label corresponding to the current feature sequence is obtained, which is the predicted bit of the transmitting end M-QAM signal.

[0065] The loss function is a BCE Loss function, i.e. a binary cross-entropy loss function, which is specifically represented as:

[0066]

[0067] Where a is the binary label 0 or 1, p(a) is the probability that the output belongs to the label, and N represents the number of groups of the predicted object.

[0068] To further verify the performance of the present invention, a system was constructed as follows: Figure 3 The experimental system shown is a 90-channel WDM-PDM-16QAM coherent optical communication system; the baud rate is 32G; the AWG operating sampling rate is 64GSa / s; the center frequency is 193.424THz, corresponding to a wavelength of 1551nm; the start and cutoff wavelengths are 1529.55nm and 1565.50nm, respectively; the channel spacing is 50GHz, covering the entire C-band; the fiber type is G652D, each fiber segment is 80km long, and a 160km loop is constructed using two 80km fiber segments and two EDFAs, and an optical circulator is used to allow the light to pass through this loop 8 times, thus transmitting a total of 1280km; an integrated coherent receiver is used to receive the signal, and a 256GSa / s oscilloscope is used to sample the signal.

[0069] Perform the following on the received signal: Figure 4 The DSP flow shown first compensates for IQ delay; then, it uses the GSOP algorithm to compensate for IQ imbalance; next, it uses the DBP algorithm to simultaneously compensate for fiber dispersion and nonlinearity, while only dispersion is compensated for signals that need to be fed into the neural network for nonlinearity compensation; then, frequency offset estimation is performed; next, the signal is matched-filtered and downsampled to twice its normal size; then, the signal is polarization demultiplexed using the CMMA algorithm; and phase noise is compensated using the BPS algorithm; then, the Volterra algorithm is used to compensate for device nonlinearity; finally, the signal with only fiber nonlinearity damage remaining uncompensated is fed into the neural network to compensate for nonlinearity.

[0070] The compensated signal is decoded and its BER (Breakpoint Error) is calculated. The Q-factor is derived from the BER and used as an indicator of the effectiveness of the invention. Figure 5 As shown, in a 90-channel 32GBaud WDM-PDM-16QAM experimental system, the CNN-BiGRU neural network based on bit-by-bit estimation outperforms neural networks with the same structure based on classification and regression.

[0071] The above experiments demonstrate that the neural network based on bit-by-bit estimation proposed in this invention can effectively improve equalization performance under the same neural network structure, and is applicable to long-distance, high-order modulation formats and multi-channel systems.

[0072] The above described embodiments are only to illustrate the preferred modes of the present application, and are not intended to limit the scope of the present application. Any modification and improvement made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A method for fiber nonlinear equalization based on a bit-wise estimation neural network, the method comprising: The method comprises the following steps: S1, obtaining M-QAM data of a sending end of a coherent optical communication system, converting the M-QAM data into bits corresponding to a modulation format of the M-QAM data, and taking the bits as neural network labels; S2, performing linear equalization on a signal transmitted through the coherent optical communication system to obtain data only retaining a nonlinear effect, and taking the data as neural network data; S3, processing the data obtained in S1 and S2, constructing a feature sequence corresponding to the data and labels, and constructing a training set and a test set based on the feature sequence and the labels; S4, constructing a CNN-BiGRU neural network based on bit-by-bit estimation, training and optimizing the neural network by using the training set in S3; S5, predicting corresponding sending end data based on the test set in S3 by using the neural network trained in S4.

2. The method of claim 1, wherein the method is based on a bit-wise estimation neural network. S1 specifically comprises: An M-QAM signal of a coherent optical communication system is acquired, and each symbol is converted into corresponding binary bits according to an encoding mode, each M-QAM signal corresponds to log2 M bits, and the bits are combined as Y i =[b1,b2,…,b log2(M) ], and Yi is a neural network label.

3. The method of claim 1, wherein, The linear equalization in S2 specifically comprises: processing the signal transmitted through the coherent optical communication system, including IQ quadrature processing, dispersion compensation, frequency offset estimation, RRC low-pass filtering, downsampling, polarization demultiplexing, and phase estimation.

4. The method of claim 1, wherein the neural network is a bit-wise estimation neural network. S3 specifically comprises: S3.1, the x, y polarization state signal x obtained in S2 is divided into real part and imaginary part and combined into one data r i , i ; considering the k data before and after, they are combined into a feature sequence X i ; i ; S3.2, using the feature sequence X i the neural network label Y obtained in S1 i Construct the training set and the test set.

5. The method of claim 4, wherein the neural network is a bit-wise estimation neural network. S3.1 specifically comprises: For a complex valued signal in the x polarization direction, it is decomposed into x i = [I x , Q x ], for a complex valued signal in the y polarization direction, it is decomposed into y i = [I y , Q y ], and then they are combined into data r i = [I x , Q x , I y , Q y ]; Consider r i The preceding k data are combined to obtain a feature sequence X i = [r i-k ,…,r i-1 ,r i ,r i+1 ,…,r i+k ].

6. The method of claim 5, wherein the neural network is a bit-wise estimation neural network. S3.2 specifically comprises: Using a sequence of features X i and a neural network label Y i Construct a training set {X i , Y i} and constructing a test set using new data according to S1-S3.

7. The method of claim 1, wherein the neural network is a bit-wise estimation neural network. S4 specifically comprises: constructing the CNN-BiGRU neural network based on bit-by-bit estimation, including an input layer, a CNN layer, a BiGRU layer, a flattening layer, a fully connected layer, a Sigmoid layer, and an output layer; The CNN-BiGRU neural network based on bit-by-bit estimation is trained using the training set {X i , Y i} based on a back propagation algorithm and an Adam optimizer by minimizing a loss function, and each parameter of the neural network is optimized.

8. The method of claim 7, wherein the neural network is a bit-wise estimation neural network. the loss function is a BCE Loss function, i.e., a binary cross-entropy loss function, and is specifically represented as: where a i is a binary label 0 or 1, p(a i ) is a probability of output belonging to a label, and N indicates a group number of a prediction object. 9.The method of claim 1, wherein, S5 specifically comprises: predicting the test set data by using the neural network trained and optimized in S4 to obtain predicted labels corresponding to the current feature sequence, i.e., bits corresponding to the predicted M-QAM signal of the sending end.

10. A fiber nonlinear equalization system based on a bit-wise estimation neural network, characterized by, The method comprises the following modules: a label acquisition module: obtaining M-QAM data of a sending end of a coherent optical communication system, converting the M-QAM data into bits corresponding to a modulation format of the M-QAM data, and taking the bits as neural network labels; a data acquisition module: performing linear equalization on a signal transmitted through the coherent optical communication system to obtain data only retaining a nonlinear effect, and taking the data as neural network data; a sequence construction module: processing the neural network data obtained by the data acquisition module to construct a feature sequence corresponding to the data; a data set construction module: constructing a training set and a test set based on the feature sequence of the sequence construction module and the neural network labels of the label acquisition module; a model construction module: constructing a CNN-BiGRU neural network based on bit-by-bit estimation; a model training module: training and optimizing the neural network by using the training set of the data set construction module; a model prediction module: predicting corresponding sending end data based on the test set of the data set construction module by using the neural network trained by the model training module.

Images