Optical fiber communication nonlinear equalization method based on early-stage monitoring and machine learning

Abstract

The invention discloses an optical fiber communication nonlinear equalization method based on early-stage monitoring and machine learning, which belongs to the field of optical fiber communication, and comprises the following steps of: before nonlinear equalization is carried out on a signal, accurately identifying the nonlinear damage degree of the signal through a monitoring module established based on machine learning; therefore, additional guidance information is provided for the nonlinear compensation module, a proper compensation model is selected from the compensation module to carry out nonlinear compensation on the signal so as to obtain an optimal compensation effect, and the nonlinear equalization performance is effectively improved. Moreover, the method only needs to output the monitoring result to the compensation module once in a short time, does not need continuous real-time monitoring, and is lower in algorithm complexity. Therefore, according to the method provided by the invention, the nonlinear equalization performance of the algorithm can be remarkably improved with relatively low algorithm complexity cost.

Description

Technical Field

[0001] The present invention belongs to the field of optical fiber communication, and more specifically, relates to an optical fiber communication nonlinear equalization method based on early monitoring and machine learning. Background Art

[0002] With the growing demand for bandwidth in the field of optical fiber communications, the development of high-speed, large-capacity, and long-distance optical fiber communication technology has become an inevitable trend in the direction of optical communications. In order to increase the transmission capacity, the development of communication systems tends to achieve higher spectral efficiency. Among them, coherent optical communication technology has the advantages of high tolerance to dispersion damage and is suitable for long-distance and large-capacity information transmission. Among them, IQ modulation methods are often used to generate high-order modulation formats. High-order modulation formats have high spectral efficiency but the Euclidean distance between constellation points is small. In order to ensure the error performance, it is necessary to increase the incident light power to improve the optical signal-to-noise ratio (OSNR), but high incident light power will cause serious nonlinear damage, which will reduce the spectral efficiency of the transmitted signal. Therefore, it is necessary to use a nonlinear equalization algorithm to compensate for nonlinear damage and improve the signal spectral efficiency.

[0003] The prior art studies the application of the combination of convolutional neural network (CNN) and bidirectional long short-term memory (biLSTM) layer in channel nonlinear equalization. Compared with the use of biLSTM layer alone, its Q factor is only improved by 0.15dB, while its algorithm complexity is improved by 94.98%. That is, although the algorithm improves the Q factor performance of the nonlinear equalization algorithm, it pays a large price in algorithm complexity and fails to achieve a good balance between algorithm complexity and equalization performance.

[0004] The prior art proposes a multiple-input multiple-output (MIMO) neural network based on a recurrent gated unit (GRU) to equalize nonlinear optical fiber damage. Compared with the biLSTM method, this method reduces the algorithm complexity by 89.8%, but at the same time the Q factor is only improved by 0.05dB. That is, although the algorithm greatly reduces the complexity of the nonlinear equalization algorithm, the Q factor performance is slightly improved, and there is still a lot of room for improvement in the equalization performance, and a good balance between the algorithm complexity and the equalization performance has not been achieved.

[0005] The prior art proposes a nonlinear equalization method and system for optical fiber communication based on convolutional neural network-bidirectional recurrent neural network (biRNN), which has a better effect in learning the nonlinear characteristics of the sequence and improves the nonlinear equalization performance. However, it does not consider whether the neural network model still maintains a high nonlinear equalization performance under different parameters such as transmitted optical power and signal transmission distance, and fails to achieve a good balance between algorithm complexity and equalization performance.

[0006] Some of the above methods improve the nonlinear equalization performance of the signal, but at the cost of a significant increase in the complexity of the algorithm; some methods reduce the complexity of the algorithm but result in poor nonlinear equalization performance. Therefore, considering the dual requirements for nonlinear equalization performance and running speed of the algorithm in actual use, how to obtain higher equalization performance with lower complexity of the equalization algorithm is an urgent problem to be solved. Summary of the invention

[0007] In response to the above defects or improvement needs of the prior art, the present invention provides a nonlinear equalization method for optical fiber communication based on preliminary monitoring and machine learning, which aims to solve the problem that the existing nonlinear equalization method cannot achieve a good balance between equalization performance and algorithm complexity, and can obtain a significant improvement in nonlinear equalization performance at the cost of lower algorithm complexity.

[0008] To achieve the above object, according to a first aspect of the present invention, a nonlinear equalization method for optical fiber communication based on early monitoring and machine learning is provided, comprising:

[0009] Training phase:

[0010] S1, taking a signal that has not been subjected to nonlinear equalization as input, taking the category of the signal as output, using a first training set to train a first target model, and using the trained target model as a monitoring module;

[0011] Wherein, the first target model is a clustering model or a neural network model; the first training set includes M types of signals that have not been subjected to nonlinear equalization, corresponding to M types of nonlinear damage degrees;

[0012] S2, taking the signal of the m-th category that has not been subjected to nonlinear equalization as input and the PRBS code corresponding thereto as output, using the second training set to train the m-th group of second target models, and using the model with the best prediction effect among the trained m-th group of second target models as the compensation model for the signal of the m-th category that has not been subjected to nonlinear equalization;

[0013] The second target model is a support vector machine or a neural network model, and the hyperparameters of each model in the mth group of second target models are different; the second training set includes M categories of signals that are not subjected to nonlinear equalization and the corresponding PRBS codes; m=1,2,…,M, M>1;

[0014] Application phase:

[0015] The signal to be nonlinearly equalized is input into the monitoring module to obtain the category of the signal to be nonlinearly equalized, and the signal to be nonlinearly equalized is input into a compensation model corresponding to its category to obtain the PRBS code of the signal to be nonlinearly equalized, thereby realizing nonlinear equalization.

[0016] According to a second aspect of the present invention, there is provided an electronic device, comprising: a computer-readable storage medium and a processor;

[0017] The computer-readable storage medium is used to store executable instructions;

[0018] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0019] According to a third aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to cause a processor to execute the method as described in the first aspect.

[0020] According to a fourth aspect of the present invention, there is provided a computer program product, comprising a computer program or instructions, wherein when the computer program or instructions are executed by a processor, the method according to the first aspect is implemented.

[0021] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0022] The method provided by the present invention accurately identifies the nonlinear damage degree of the signal through a monitoring module established based on machine learning before performing nonlinear equalization on the signal, thereby providing additional guidance information for the nonlinear compensation module, and then selecting a suitable compensation model in the compensation module to perform nonlinear compensation on the signal to obtain the best compensation effect, thereby effectively improving the nonlinear equalization performance; and the method only needs to output the monitoring result to the compensation module once in a short time, without the need for continuous real-time monitoring, and has a lower algorithm complexity. Therefore, the method proposed by the present invention can significantly improve the algorithm nonlinear equalization performance at the cost of lower algorithm complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of a nonlinear equalization method for optical fiber communication based on preliminary monitoring and machine learning provided by an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of the CNN model structure provided by an embodiment of the present invention;

[0025] Figure 3 A schematic diagram of the ANN model structure provided by an embodiment of the present invention;

[0026] Figure 4 It is a schematic diagram of a signal generating device of a coherent optical communication system;

[0027] Figure 5 A schematic diagram of the structure of a monitoring model based on CNN provided in an embodiment of the present invention;

[0028] Figure 6 A schematic diagram of the structure of a nonlinear compensation model based on ANN provided in an embodiment of the present invention;

[0029] Figure 7 A flow chart of the training phase in the optical fiber communication nonlinear equalization method based on early monitoring and machine learning provided by an embodiment of the present invention;

[0030] Figure 8 A flow chart of the application phase of the optical fiber communication nonlinear equalization method based on preliminary monitoring and machine learning provided by an embodiment of the present invention;

[0031] Fig. 9 This is a diagram of the preliminary monitoring results of the nonlinear damage parameters of the signal to be tested provided by an embodiment of the present invention (taking the transmitted optical power and signal transmission distance as examples).

[0032] Fig.10 A performance comparison diagram of a signal to be processed provided by an embodiment of the present invention, after nonlinear compensation without using preliminary monitoring and after nonlinear compensation using preliminary monitoring proposed by the present invention;

[0033] Fig.11 A comparison diagram of the complexity of the nonlinear equalization algorithm provided by the embodiment of the present invention for different tag signals without using a monitoring module and with using the monitoring module as an aid;

[0034] In all drawings, the same reference numerals are used to denote the same elements or structures, wherein:

[0035] 1-main optical signal transmitter, 2-bypass optical signal transmitter, 3-wavelength division multiplexer, 4-standard single-mode optical fiber, 5-erbium-doped fiber amplifier, 6-wavelength division multiplexer, 7-signal receiving device. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0037] The embodiment of the present invention provides a nonlinear equalization method for optical fiber communication based on early monitoring and machine learning, such as Figure 1 As shown, including:

[0038] Training phase:

[0039] S1, taking a signal that has not been subjected to nonlinear equalization as input, taking the category of the signal as output, using a first training set to train a first target model, and using the trained target model as a monitoring module;

[0040] Wherein, the first target model is a clustering model or a neural network model; the first training set includes M different types of signals without nonlinear equalization, corresponding to M different degrees of nonlinear damage;

[0041] S2, taking the signal of the mth category without nonlinear equalization as input, taking the PRBS code of the signal of the mth category without nonlinear equalization as output, using the second training set to train the mth group of second target models, and taking the model with the best prediction effect among the trained mth group of second target models as the compensation model of the signal of the mth category without nonlinear equalization, thereby forming a compensation module;

[0042] The second target model is a neural network model, and the hyperparameters of each model in the mth group of second target models are different; the second training set includes M types of signals that are not subjected to nonlinear equalization and the corresponding PRBS codes; n=1, 2, ..., M, M>1;

[0043] Application phase:

[0044] The signal to be nonlinearly equalized is input into the monitoring module to obtain the category of the signal to be nonlinearly equalized, and the signal to be nonlinearly equalized is input into the compensation model corresponding to its category for nonlinear compensation to obtain the PRBS code of the signal to be nonlinearly equalized, thereby realizing nonlinear equalization.

[0045] Specifically, the input of the monitoring module, that is, the signal to be nonlinearly equalized, can be time domain data such as signal constellation diagram, eye diagram, etc., or frequency domain data such as signal RF spectrum, signal spectrum, etc., or polarization domain data such as Stokes parameters, etc.

[0046] The signal to be nonlinearly equalized is signal data received by a coherent optical communication receiver and subjected to preliminary digital signal processing. Preferably, signal data subjected to linear equalization processing such as coherent reception, analog-to-digital conversion, dispersion equalization, and channel demultiplexing is used as the signal to be nonlinearly equalized.

[0047] Considering the effect of nonlinear equalization and the complexity of data processing, preferably, the data to be equalized (ie, the signal to be nonlinearly equalized) is time domain data.

[0048] The monitoring module is a machine learning model trained in the training phase (i.e., the first target model), which is used to obtain parameters related to the degree of nonlinear damage suffered by the signal in the application phase, and then determine the signal parameter settings that affect the degree of nonlinear damage, thereby distinguishing the signal according to different categories.

[0049] The above-mentioned parameter related to the degree of nonlinear damage suffered by the signal can be selected from at least one of the incident light power, signal transmission distance, number of channels, channel spacing and optical fiber nonlinear coefficient.

[0050] The first target model can be a clustering model belonging to unsupervised learning, such as an autoencoder, a k-nearest neighbor model, etc.; or a neural network model belonging to supervised learning, such as an artificial neural network (ANN), CNN, recurrent neural network (RNN) and deep neural network (DNN), etc.

[0051] Considering that the convolutional neural network has a strong feature extraction capability and can improve the accuracy of monitoring, preferably, the first target model is CNN. As an example, Figure 2 As shown, CNN includes:

[0052] There are i convolutional layers with j channels. 2i-1 , the length of a single channel is n i , the convolution kernel size is p i ,This layer uses convolution kernels to process RF spectrum data and extract local features of spectrum data.

[0053] There are k pooling layers in total, with j channels 2k , the length of a single channel is m k , the pooling kernel size is q k ,This layer performs maximum pooling on the data, retaining the extracted data features while reducing the amount of data.

[0054] The flattening layer is used to flatten the multi-channel data in the previous pooling layer into one-dimensional data.

[0055] The hidden layer, hidden layer and flat layer are fully connected after the activation function to transfer data.

[0056] Output layer: The monitoring results output by the output layer are parameters that characterize the degree of nonlinear damage of the signal, corresponding to the signal category.

[0057] The CNN model is used to monitor the types of signals in the signal sample library, such as the optical power emitted by different signals, signal transmission distance, number of channels, channel spacing, modulation format, signal rate, and fiber nonlinear parameters, and use the single label N output by the model to distinguish the signal categories.

[0058] For example, if the transmit optical power and signal transmission distance are used to ensure the nonlinear damage degree of the signal, the label N is used to represent the signal type of different transmit optical power and signal transmission distance combinations.

[0059] According to the label N output by the CNN model, it is used to guide the nonlinear compensation module to select a model with the best equalization performance from the second target model group corresponding to the signal category.

[0060] That is, in the application stage, the signal to be nonlinearly equalized is input into the monitoring module to monitor the degree of nonlinear damage of the signal (characterized by the type of signal). The monitoring module outputs the category of the signal to be nonlinearly equalized, which is used to guide the nonlinear compensation module to select a model with the best equalization performance from the second target model group corresponding to the signal category, and input the signal to be nonlinearly equalized into the model with the best equalization performance to perform nonlinear compensation on it.

[0061] The second target model can be a support vector machine or a neural network model belonging to supervised learning, such as ANN, CNN, RNN, DNN and biLSTM neural network.

[0062] During the training phase, each category of signals to be nonlinearly equalized corresponds to a set of second target models, and each model in the set of second target models has different hyperparameters.

[0063] Preferably, the hyperparameters include at least one of the symbol window length of the input layer of the neural network model, the number of hidden layers, the length of hidden layer neurons, the number of hidden layer neurons, the loss function, the optimization function, and the regularization parameter.

[0064] Considering that the structure of ANN is simple and can reduce computational complexity, preferably, the second target model is ANN, and the hyperparameters include the symbol window length of the input layer, the number of hidden layers and the length of single-layer neurons.

[0065] As an example, Figure 3 As shown, ANN includes:

[0066] Input layer: The length of this layer directly affects the length of the time domain symbol data input to the ANN. The length of different symbol windows will affect the ability of the ANN model to compensate for the inter-symbol interference caused by nonlinear damage. Therefore, in the training stage, for the signals to be equalized with different input optical power and signal transmission distance, it is necessary to set different input layer symbol window lengths for the mth group of ANN models corresponding to the mth category for training.

[0067] Hidden layer: Since different numbers of hidden layers and lengths of hidden layer neurons will significantly affect the nonlinear mapping ability of the ANN model, the hidden layer parameters will also affect the effect of compensating for nonlinear damage. For signals to be equalized with different input optical powers and signal transmission distances, it is necessary to set different numbers of hidden layers and lengths of hidden layer neurons for training the mth group of ANN models corresponding to the mth category.

[0068] At the output layer, the ANN model outputs the time domain symbol data after nonlinear equalization.

[0069] The method provided by the present invention is described below with a specific example.

[0070] Take the monitoring module based on CNN and the nonlinear compensation module based on ANN as an example, and the signal to be nonlinearly equalized is a time domain signal. The time domain signal to be nonlinearly equalized is used as the input of the entire method, and then the signal is FFTed to convert the time domain signal into a radio frequency spectrum as the input of the monitoring module. In this example, in order not to introduce an additional FFT step in the method to increase the complexity of the algorithm, the frequency domain signal data in the dispersion frequency domain compensation in the linear equalization process is directly used as the input of the monitoring module. Then the CNN model performs the classification task, identifies the incident optical power and signal transmission distance parameters corresponding to the frequency domain data, and outputs the label N to represent the signal category corresponding to the signal parameters and serves as the output of the monitoring module. In the compensation module, the time domain symbol is directly used as the input of the ANN model, and the output label N of the monitoring module is used to guide the compensation module to select the ANN model corresponding to the symbol window length and neural network parameters. Then the ANN performs the regression task, outputs the nonlinear residual value, and then the input time domain symbol and the output residual value are subtracted to obtain the time domain symbol after nonlinear equalization.

[0071] Since nonlinear damage will broaden the signal spectrum and cause an asymmetric oscillation multi-peak structure, and the degree of broadening and oscillation will increase with the increase of nonlinear damage, the CNN in the monitoring module can capture the local characteristics of the RF spectrum, so it has a high monitoring accuracy and can accurately distinguish optical signals with different transmission optical powers and signal transmission distances.

[0072] Similarly, nonlinear damage in the time domain will cause inter-code crosstalk in the time domain symbol data, thus distorting the time domain signal, and the degree of distortion will also increase with the increase of the degree of nonlinear damage. The ANN algorithm has a strong nonlinear mapping capability, but the impact of different degrees of nonlinear damage on the signal is different. At this time, for time domain signals with different degrees of nonlinear damage, the symbol window length and neural network parameters with the best nonlinear equalization effect are also different. Therefore, it is necessary to train ANN models with different parameters, and then use the label N output by the monitoring module to guide the compensation module to select the ANN model with the best nonlinear equalization effect, so as to obtain better nonlinear equalization performance.

[0073] In order to verify and optimize the effect of the model, this embodiment uses Figure 4 The coherent optical communication signal generating device shown obtains signal samples affected by nonlinear damage for model training. Its structure includes: 1-main optical signal transmitter, 2-bypass optical signal transmitter, 3-wavelength division multiplexer, 4-standard single-mode optical fiber, 5-erbium-doped fiber amplifier, 6-wavelength division multiplexer, and 7-signal receiving device. The signal simulation structure can simulate the generation of optical signals in a real optical communication system. The optical transmitter generates optical signals with modulation parameters such as specific modulation format, bit rate, pulse shape, number of channels and channel spacing, and simulates different degrees of nonlinear damage by changing the incident optical power and signal transmission distance of the optical transmitter. After wavelength division multiplexing, the signal is transmitted to the optical fiber span, and the erbium-doped fiber amplifier is used to compensate for the transmission loss. After being transmitted through several optical fiber spans, the signal passes through the wavelength division multiplexer to extract the main channel separately, and finally is received by the coherent optical receiver and analog-to-digital conversion is performed to obtain electrical time domain symbol data samples with different parameter settings. It should be noted that, in addition to the two types of nonlinear damage in this embodiment, other nonlinear damage factors in the signal simulation structure may also be adjusted to simulate and obtain signal samples carrying other nonlinear damage information.

[0074] This embodiment selects various parameters widely used in optical fiber communication systems after comprehensive consideration: for example, the signal modulation format is 16QAM; the signal rate is 28Gbaud; the pulse shaping method adopts Nyquist; the number of wavelength division multiplexing channels is three channels, and the channel interval is 50Gbaud.

[0075] Considering that the nonlinear equalization method provided by the present invention targets nonlinear damage that affects the signal quality in the optical fiber, the simulation system is used to verify and analyze the signals of nonlinear damage factors such as different incident optical powers and signal transmission distances. Finally, in this embodiment, the parameter settings of incident optical power of 12-16dBm, step size of 2dBm, signal transmission distance of 8-14 spans, step size of 2 spans, and each span of 80km are adopted to construct a signal data sample set under high incident optical power and long signal transmission distance.

[0076] According to the signal parameters and nonlinear damage factor settings of this embodiment, taking the nonlinear damage degree characterized by incident optical power and signal transmission distance as an example, 12 signal types with different incident optical power and signal transmission distance combinations are set, and 800 groups of different signal samples are collected for each signal. These signals are assigned labels 0-11 respectively, and the total number of signal samples is 9600 groups. The signal sample library is used for CNN model training and testing in the monitoring module. At the same time, the time domain symbol data sample library used for ANN model training and testing in the nonlinear compensation module in this embodiment is a symbol length of 2 20 The time domain signal data corresponding to the pseudo-random binary sequence (PRBS) code with a known sequence as the label is used for training and testing the ANN model. The PRBS code order used in this embodiment is a longer 31 order to prevent the ANN from learning the hidden rules of the PRBS code during the training process.

[0077] In this embodiment, the structure of CNN in the monitoring module is as follows: Figure 5As shown, it includes: an input layer, two convolution layers, two pooling layers, a flat layer, two hidden layers, and an output layer; specifically, the first layer is the input layer, and the one-dimensional RF spectrum signal data is used as the input of the CNN model, with a length of 8192 and a covered spectrum bandwidth of 50GHz; the second layer is the convolution layer 1, with 2 channels, a convolution kernel size of 1×5, and a step size of 2, so the single layer length is 4096; the third layer is the pooling layer 1, with 2 channels, a pooling kernel size of 1×2, and a step size of 2, so the single layer length is 2048; the fourth layer is the convolution layer 2, with 4 channels, a convolution kernel size of 1×5, and a step size of 2, so the single layer length is 1024; The fifth layer is pooling layer 2, with 4 channels, a pooling kernel size of 1×2, and a stride of 2, so the length of a single layer is 512; the sixth layer is a flattening layer, which flattens the previous multi-channel pooling layer into one-dimensional data, so the size of this layer is 1×2048; the seventh layer is hidden layer 1, which is a fully connected layer, fully connected to the previous flattening layer, and has a size of 1×32; the eighth layer is hidden layer 2, which is a fully connected layer, fully connected to the previous hidden layer 1, and has a size of 1×32; the ninth layer is the output layer, which is a fully connected layer, fully connected to the previous hidden layer 2, corresponding to 12 different signal labels of incident optical power and signal transmission distance, so the size is 1×12. Optionally, between the convolution layer and the pooling layer, as well as between the flattening layer and the hidden layer, and between the hidden layer and the hidden layer, the ReLu function is selected as the activation function, and the Adam optimizer and the cross entropy loss function are selected, the learning rate is set to 1e-3, and the learning rate decay factor is set to 1e-3.

[0078] In this embodiment, the structure of the ANN in the nonlinear compensation module is as follows: Figure 6 As shown, it includes: an input layer, two hidden layers, and an output layer; specifically, the first layer is the input layer, and the one-dimensional time domain symbol data is used as the input of the ANN model. The sampling rate of the time domain symbol data is twice the baud rate of the transmitted signal, the input symbol window length is 5-29, and the step size is 4. The subsequent hidden layer is fully connected to the input layer, the number of hidden layers is one or two, and the length of neurons in each layer is 5, 10, 20, 40, etc. By combining different numbers of hidden layers with single-layer neuron lengths, this embodiment combines 8 different neural network structure parameter settings (that is, for a signal category, a group of ANN models trained include 8 ANN models with different neural network structure parameters). The output layer is a single neuron, and is fully connected to the previous hidden layer. Since the ANN performs a regression task at this time, the output value is the nonlinear residual value of the time domain symbol. Optionally, the ReLu function is selected as the activation function between the input layer and the hidden layer, and between the hidden layer and the hidden layer. The Adam optimizer and the mean square error (MSE) loss function are selected at the same time. The learning rate is set to 1e-3, and the learning rate decay factor is set to 1e-3.

[0079] In this embodiment, the algorithm flow of the training phase of the monitoring module and the nonlinear compensation module is as follows: Figure 7 As shown in the figure. Since the number and length of the convolutional layer and pooling layer of the model will affect the ability of CNN to extract local features of the spectrum, and thus affect the monitoring accuracy, a large number of samples are required for training. The 9600 sets of signal samples with different incident optical power and signal transmission distance collected are divided into training sample sets and test sample sets in a ratio of 7:3. The early stopping strategy is used in the CNN training process to stop the training after the model test loss value in the training process no longer decreases in a certain number of training rounds, so as to prevent the CNN model from overfitting. Finally, after 300 epochs of training, the model converges and reaches stability.

[0080] For the ANN model of the nonlinear compensation module of this embodiment, the signal samples are divided into sample libraries corresponding to different labels N according to different incident optical powers and signal transmission distances. The data length of the time domain symbol sample library for each label is 2 20 , each time a certain symbol window length of data is intercepted from the sample library as a set of data to enter the ANN input layer, the data length is 5-29, the compensation is 4, that is, the same size as the ANN model input layer. During the training process, the optuna optimization algorithm is used to optimize the ANN model symbol window length and neural network structure parameters corresponding to each signal label N, where the optimization range of the symbol window length is 5-29, and the optimization range of the hidden layer is one to two layers. The sample label of the training set is the difference between the time domain symbol data under ideal conditions and the time domain symbol data actually input into the neural network. After the training is completed, the optimal symbol window length and neural network structure parameters under the signal corresponding to each label N are obtained.

[0081] In this embodiment, the algorithm flow of the test phase of the monitoring module and the nonlinear compensation module is as follows: Figure 8 As shown. During the test, the time domain symbol data to be monitored is processed by FFT to generate RF spectrum data, which is then input into CNN for testing, and the label N representing the incident optical power of the signal and the signal transmission distance is output. In the nonlinear compensation module, the appropriate ANN model is selected for the compensation module according to the label N output by the monitoring module, that is, the symbol window length and neural network structure parameters corresponding to the label N signal. Finally, the nonlinear residual value is output to obtain the time domain symbol data after nonlinear equalization.

[0082] In this embodiment, the early monitoring results of the signal nonlinear damage parameters are as follows: Fig. 9As shown. For 28Gbaud-PDM-16QAM signals, the average recognition accuracy of the signal label type by the monitoring module can reach 99.86%, among which only the signal label with incident light power of 16dBm and signal transmission distance of 1120km has a monitoring accuracy of 98.4%, and the accuracy of other labels is 100%. Therefore, it can be concluded from this embodiment that the overall accuracy of the monitoring module is relatively high, which meets the requirements of the subsequent nonlinear compensation module for monitoring performance. At the same time, the law of monitoring accuracy is also consistent with the theory, that is, when the incident light power is large and the signal transmission distance is long, the nonlinear damage to the signal is large, and the local detail characteristics of the RF spectrum are seriously distorted, resulting in a decrease in monitoring accuracy.

[0083] In this embodiment, the nonlinear equalization performance comparison of the nonlinear equalization under different incident optical powers and signal transmission distances is shown in FIG. Fig.10 As shown. At this time, the signal transmission power is 14dBm, and the law of the change of Q factor equalization performance with the transmission optical power and signal transmission distance is also consistent with the theory. Under the same incident optical power, as the signal transmission distance increases, the signal Q factor gradually decreases; under the same signal transmission distance, as the incident optical power increases, the signal Q factor gradually decreases. In this embodiment, after using nonlinear equalization that does not include preliminary monitoring, compared with using only linear equalization, the signal Q factor is increased by an average of 0.89dB; and the use of nonlinear equalization that includes preliminary monitoring is compared with using only linear equalization. Compared with using only linear equalization, the signal Q factor is increased by an average of 1.09dB. That is, after using the CNN-based monitoring module in this embodiment, the Q factor of the nonlinear equalization is increased by an average of 0.2dB.

[0084] At the same time, in this embodiment, the impact of preliminary monitoring on the complexity of the algorithm is considered, and the complexity of the algorithm is mainly affected by the number of multiplications in the algorithm. Since the multiplication operations of the CNN model in this embodiment are mainly concentrated in the convolution kernel operations of the CNN convolution layer and the fully connected parts between the two layers of the neural network, the pooling layer only performs the maximum pooling operation, and the complexity of the pooling layer can be ignored. Therefore, in the nonlinear equalization method based on preliminary monitoring and machine learning, the CNN-based monitoring module introduces additional multiplications. In the CNN model of this embodiment, the length of each convolution layer is n. i , the number of channels is j i , the size of each convolution kernel is p i (i=1, 2); followed by a sequence of length j 2 m 2 A flat layer with length n 3 The hidden layer has a length of n 4The output layer, hidden layer and flat layer are fully connected after activation function, so the number of multiplications in the full connection process will double. At this time, the algorithm complexity of the single monitoring module can be expressed as formula (1):

[0085]

[0086] The multiplication operation of the ANN model in this embodiment is mainly concentrated in the fully connected part between the input layer and the hidden layer, and between the hidden layer and the hidden layer. Therefore, in the nonlinear equalization method based on early monitoring and machine learning, the nonlinear compensation module based on ANN contains the basic number of multiplications. In the ANN model of this embodiment, the length of each input layer, hidden layer and output layer is k respectively. i (i=1, 2, 3, 4), the output layer and the hidden layer, and the hidden layer and the hidden layer are all fully connected after the activation function. At this time, the algorithm complexity of the single monitoring module can be expressed as formula (2):

[0087] C ANN =2(k 1 k 2 +k 2 k 3 +k 3 k 4 ) (2)

[0088] In this embodiment, the complexity comparison of the nonlinear equalization algorithm without using the monitoring module and with using the monitoring module for different tag signals is as follows: Fig.11 As shown. Since the time domain symbol sample size of a single type of signal label in the nonlinear compensation module is 2 20 Therefore, when calculating the number of multiplications consumed by the nonlinear equalization method to restore a single symbol, it is necessary to divide the number of multiplications of the monitoring module by the total number of time domain symbols that the nonlinear compensation module needs to restore. In this embodiment, it can be calculated from equations (1) and (2) that after adding the monitoring module, 20 more multiplications are required to restore a single time domain symbol, which increases the algorithm complexity of signals of different label categories by an average of 0.25%.

[0089] An embodiment of the present invention provides an electronic device, including: a computer-readable storage medium and a processor;

[0090] The computer-readable storage medium is used to store executable instructions;

[0091] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method described in any of the above embodiments.

[0092] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method described in any of the above embodiments.

[0093] An embodiment of the present invention provides a computer program product, including a computer program or instructions, and when the computer program or instructions are executed by a processor, the method described in any of the above embodiments is implemented.

[0094] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A nonlinear equalization method for optical fiber communication based on early monitoring and machine learning, characterized in that: include: Training phase: S1, taking a signal that has not been subjected to nonlinear equalization as input, taking the category of the signal as output, using a first training set to train a first target model, and using the trained target model as a monitoring module; Wherein, the first target model is a clustering model or a neural network model; the first training set includes M types of signals that have not been subjected to nonlinear equalization, corresponding to M types of nonlinear damage degrees; S2, taking the signal of the m-th category that has not been subjected to nonlinear equalization as input and the PRBS code corresponding thereto as output, using the second training set to train the m-th group of second target models, and using the model with the best prediction effect among the trained m-th group of second target models as the compensation model for the signal of the m-th category that has not been subjected to nonlinear equalization; The second target model is a support vector machine or a neural network model, and the hyperparameters of each model in the mth group of second target models are different; the second training set includes M categories of signals that are not subjected to nonlinear equalization and the corresponding PRBS codes; m=1,2,…,M, M>1; Application phase: The signal to be nonlinearly equalized is input into the monitoring module to obtain the category of the signal to be nonlinearly equalized, and the signal to be nonlinearly equalized is input into a compensation model corresponding to its category to obtain the PRBS code of the signal to be nonlinearly equalized, thereby realizing nonlinear equalization.

2. The method according to claim 1, characterized in that The signal to be nonlinearly equalized is time domain data, frequency domain data or polarization domain data.

3. The method according to claim 1 or 2, characterized in that The nonlinear damage degree is characterized by at least one of incident optical power, signal transmission distance, number of channels, channel spacing or optical fiber nonlinear coefficient.

4. The method according to claim 1, characterized in that The first target model is CNN.

5. The method according to claim 1, characterized in that The hyperparameters include at least one of the symbol window length of the input layer of the neural network model, the number of hidden layers, the length of hidden layer neurons, the number of hidden layer neurons, the loss function, the optimization function, and the regularization parameter.

6. The method according to claim 1 or 5, characterized in that The second target model is ANN, and the hyperparameters include the symbol window length of the input layer, the number of hidden layers and the length of single-layer neurons.

7. An electronic device, characterized in that: include: A computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method according to any one of claims 1 to 6.

9. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images