Coherent light nonlinear IQ damage equalization method and device

Abstract

The invention discloses a coherent light nonlinear IQ damage equalization method and device. The method comprises the following steps: obtaining 16 QAM data to be equalized; preprocessing the 16QAM data to obtain preprocessed real part data and imaginary part data; inputting the preprocessed real part data into a pre-trained first SD-ResNet model to obtain balanced real part data; inputting the preprocessed imaginary part data into a pre-trained second SD-ResNet model to obtain balanced imaginary part data; and obtaining the equalized 16QAM data according to the equalized real part data and the equalized imaginary part data, so that the expression capability of nonlinear mapping is ensured while the calculation complexity is reduced, and the recovery precision of the IQ data is improved.

Description

Technical Field

[0001] The present application belongs to the field of optical transmission communication technology, and specifically relates to a method and device for coherent optical nonlinear IQ damage equalization. Background Art

[0002] In recent years, with the rapid development of Internet technology and the advent of the big data era, people's demand for high-speed and large-capacity communication networks has continued to increase. Traditional telecommunications networks can no longer meet such huge data traffic needs, and optical fiber communication networks have become the best choice due to their excellent broadband characteristics. However, the bandwidth resources of optical fiber communication networks are not unlimited. How to achieve higher information transmission rates on limited optical fiber resources has become a key issue that needs to be urgently solved in the field of optical communications. Coherent optical communication systems combine advanced modulation technologies such as high-order quadrature amplitude phase modulation (QAM) to achieve more efficient information transmission than traditional optical communications, thus becoming an important means to improve the spectrum efficiency of optical fiber networks. High-order QAM modulation greatly improves the effective bandwidth utilization of optical fibers by cleverly combining and encoding the amplitude and phase of the carrier. At present, coherent optical communication systems can support 120 GBaud QAM modulation, and each symbol can transmit more than 10 bits of information, becoming a key technology to significantly improve the total throughput of optical fiber networks. However, while high-order QAM modulation promotes the development of high symbol rates, the equalization technology of coherent optical communication systems faces many challenges. High-order modulation formats are more susceptible to damage to optical components. Imperfections such as phase noise of the transmitting laser and local oscillator, fiber nonlinearity, and imbalance between the IQ channels of the coherent optical receiver front end severely affect the performance of the QAM format. In particular, imperfections in the receiver front end can have a considerable impact on the transmission performance. Transceivers using large-order quadrature and amplitude modulation (QAM) formats usually require high signal-to-noise ratio values ​​for reliable transmission, which makes their implementations less tolerant to imbalances in the receiver front end. Therefore, two important questions in this context are how these imperfections relate to the performance of signal detection and equalization, and how to design the DSP to mitigate the resulting penalty.

[0003] Recent studies have proposed some architectures that are robust to receiver front-end imperfections to improve the system's tolerance to IQ impairments. Constant modulus algorithm (CMA), as the most commonly used equalization algorithm in coherent optical communications, iteratively adjusts the coefficients of the equalization filter to keep the amplitude of the filter output symbol consistent with a known constant value. For the problems existing in standard CMA, asymmetric CMA can be used to eliminate polarization interference. In addition, for polarization mode dispersion (PMD) caused by differential signal polarization propagation, adaptive CMA algorithms have been proposed for PMD compensation. These algorithms usually presuppose the size of the filter, which may impair the filter performance. Using the concept of pseudo-fractional tap length (FT), the weights and tap lengths of the optimal equalizer can be estimated simultaneously, thereby enhancing the equalization process with higher efficiency in blind mode. In the case of IQ deviation in the transmitter, a 6×6 multiple-input multiple-output (MIMO) structure is used, and the 𝑁1 tap is used for polarization recovery and receiver IQ deviation calibration. Aiming at the frequency domain nonlinear effect of dispersion-enhanced spectral aliasing under symbol rate sampling conditions, a hardware-efficient and dispersion-tolerant equalization scheme is proposed. Compared with T / 2-spaced 2×2 MIMO, it simplifies the equalization structure and saves about 50% of real-valued multipliers. However, it should be noted that the above-mentioned traditional digital equalizer methods mainly use deterministic models and linear / nonlinear algorithms, both of which require pre-knowledge of carrier information before the channel equalization process, which may consume a lot of overhead to achieve accurate estimation. It is used to compensate for signal distortion caused by various physical effects in coherent optical communication systems. However, there are more complex non-ideal effects in actual optical fiber links, such as laser phase noise, polarization effect, filtering effect, etc., which make the model deviate from the actual situation, thereby affecting the compensation effect of IQ damage of coherent receivers. Therefore, the traditional equalization algorithm cannot use prior information, and it becomes more difficult to accurately restore the original IQ information, which is difficult to meet the requirements of coherent optical communication systems for high-order QAM formats and high symbol rates. Summary of the invention

[0004] Purpose: In view of at least one of the above technical problems, the present application provides a coherent optical nonlinear IQ damage equalization method and device.

[0005] The technical solution adopted in this application is:

[0006] In a first aspect, the present application provides a coherent optical nonlinear IQ impairment equalization method, comprising:

[0007] S1. Obtain 16QAM data to be equalized;

[0008] S2, preprocessing the 16QAM data to obtain preprocessed real data and imaginary data;

[0009] S3, input the preprocessed real data into the pretrained first SD-ResNet model to obtain balanced real data; input the preprocessed imaginary data into the pretrained second SD-ResNet model to obtain balanced imaginary data;

[0010] S4, obtaining equalized 16QAM data according to the equalized real part data and the equalized imaginary part data;

[0011] The method for obtaining the pre-trained first SD-ResNet model and the second SD-ResNet model includes:

[0012] The network structures of the first SD-ResNet model and the second SD-ResNet model are the same, and both include: an input layer, four residual layers and an output layer; the input layer includes a 1-dimensional convolution layer, a batch normalization layer and a ReLU activation function; each residual layer includes two residual blocks, and each residual block includes a 1-dimensional convolution layer, a batch normalization layer and a ReLU activation function; the output layer includes a global average pooling layer, a flattening layer and a fully connected layer;

[0013] The first SD-ResNet model is trained using a labeled real data sample set, and the second SD-ResNet model is trained using a labeled imaginary data sample set until the preset conditions are met to obtain the trained first SD-ResNet model and the second SD-ResNet model; during the training process, for each residual block, a decision is made whether to skip it based on a random depth mechanism. If the randomly generated number is less than the survival probability of the residual block, the calculation of the residual block is skipped and the input is directly returned; if the randomly generated number is greater than or equal to the survival probability of the residual block, the residual block is calculated.

[0014] In a second aspect, the present application provides a coherent optical nonlinear IQ impairment equalization device, including a processor and a storage medium;

[0015] The storage medium is used to store instructions;

[0016] The processor is configured to operate according to the instructions to execute the method according to the first aspect.

[0017] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the method described in the first aspect when executed by a processor.

[0018] In a fourth aspect, the present application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the method of the first aspect when executing the computer program.

[0019] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which implements the method described in the first aspect when executed by a processor.

[0020] Beneficial effects: The coherent optical nonlinear IQ damage equalization method and device provided in the present application have the following advantages: In the present application, a ResNet IQ data equalization scheme based on stochastic depth (Stochastic Depth) is proposed. The IQ information of the transmitted symbol and the received symbol is used as the input and output of ResNet, and the residual idea is innovatively used to connect the input and output information. The IQ data is standardized by Z-score, and the SD mechanism is introduced in ResNet to control the number of network modules involved in the training. SD-ResNet learns the IQ nonlinear offset mapping from input to output from the original data end-to-end, without the need for an explicit modeling process, and discovers the implicit IQ data rules and characteristic patterns. In addition, in order to prevent redundancy in network training, a Stochastic Depth (SD) mechanism is added to the network. While ensuring the expression ability of the network's nonlinear mapping, the computational complexity of the SD-ResNet network model is reduced, and the recovery accuracy of the IQ data is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a flow chart of a method for equalizing coherent optical nonlinear IQ impairments according to an embodiment of the present application;

[0022] Figure 2 Schematic diagram of the SD-ResNet network structure according to an embodiment of the present application;

[0023] Figure 3 Schematic diagram of the SD ResNet random depth mechanism according to an embodiment of the present application;

[0024] Figure 4 A schematic diagram of loss function values ​​corresponding to different survival probabilities according to an embodiment of the present application;

[0025] Figure 5 A schematic diagram of k cross validation according to an embodiment of the present application;

[0026] Figure 6 Schematic diagram of SD-ResNet performance under different training batch sizes according to an embodiment of the present application;

[0027] Figure 7 A schematic diagram of the accuracy of the k-cross-validation SD-ResNet model according to an embodiment of the present application;

[0028] Figure 8 It is a schematic diagram of bit error rate performance according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] The present application is further described below in conjunction with the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present application, and cannot be used to limit the protection scope of the present application.

[0030] In the description of this application, "several" means more than one, "more" means more than two, "greater than", "less than", "exceed", etc. are understood to exclude the number itself, and "above", "below", "within", etc. are understood to include the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0031] In the description of the present application, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0032] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0033] Embodiment 1: This embodiment provides a coherent optical nonlinear IQ damage equalization method, such as Figure 1 As shown, including:

[0034] S1. Obtain 16QAM data to be equalized;

[0035] S2, preprocessing the 16QAM data to obtain preprocessed real data and imaginary data;

[0036] In some embodiments, step S2 includes:

[0037] Performing Z-standardization processing on the real data and the imaginary data of the 16QAM data respectively to obtain standardized real data and imaginary data;

[0038] The standardized real data and imaginary data are respectively dimensionally expanded (from two-dimensional data to three-dimensional data) to obtain preprocessed real data and imaginary data.

[0039] In this embodiment, if Figure 1 As shown, the real data and imaginary data of the 16QAM data to be equalized are separated into X and X', respectively, and after Z standardization is performed on them, the standardized real data x are obtained respectively. i-norm , imaginary data and x ’ i-norm In order to adapt the input dimension of the SD-ResNet model, x i-norm and x ’ i-norm Perform dimension expansion, expanding the original two-dimensional data into three-dimensional data.

[0040] S3, input the preprocessed real data into the pretrained first SD-ResNet model to obtain balanced real data; input the preprocessed imaginary data into the pretrained second SD-ResNet model to obtain balanced imaginary data;

[0041] The method for obtaining the pre-trained first SD-ResNet model and the second SD-ResNet model includes:

[0042] The network structures of the first SD-ResNet model and the second SD-ResNet model are the same, both of which include: an input layer, four residual layers and an output layer; Figure 2 As shown in the figure, the input layer includes a 1D convolution layer, a batch normalization layer, and a ReLU activation function; each residual layer includes two residual blocks, and each residual block includes a 1D convolution layer, a batch normalization layer, and a ReLU activation function; the output layer includes a global average pooling layer, a flattening layer, and a fully connected layer;

[0043] The first SD-ResNet model is trained using a labeled real data sample set, and the second SD-ResNet model is trained using a labeled imaginary data sample set until the preset conditions are met to obtain the trained first SD-ResNet model and the second SD-ResNet model; during the training process, for each residual block, a decision is made whether to skip it based on a random depth mechanism. If the randomly generated number is less than the survival probability of the residual block, the calculation of the residual block is skipped and the input is directly returned; if the randomly generated number is greater than or equal to the survival probability of the residual block, the residual block is calculated.

[0044] In some embodiments, the survival probability of the residual blocks is 0.7.

[0045] In some embodiments, to obtain x i-norm Take the training task of the first SD-ResNet model with real data features as an example.

[0046] First, the real data x after dimension expansion i-norm The input layer of the SD-ResNet model includes a 1D convolution layer (Conv1d), a batch normalization layer (BatchNorm1d), and a ReLU activation function. The data is downsampled through the maximum pooling layer to reduce data redundancy.

[0047] The data is downsampled and enters the residual layer. The residual layer has four layers, each containing two residual blocks. After the data to be processed goes through these steps, the survival probability generated by the random depth mechanism determines whether the residual block needs to be skipped. Through experiments, it is found that the survival probability of each residual block is 0.7. After the residual layer is processed, the feature vector is obtained, and the feature vector enters the output layer. The output layer contains a global average pooling layer, a flattening layer, and a fully connected layer. The global average pooling layer is used for adaptive average pooling to reduce the data to a fixed size. The flattening layer flattens the data into a one-dimensional vector. Finally, the fully connected layer maps the data to the number of output categories and outputs the classification results of the data. The following is a detailed description of the entire training process.

[0048] x i is the i-th real data input sample, y i is the true output of the i-th input sample, f(xi;θ) is the predicted output of the SD-ResNet model, and the connected convolutional layer, batch normalization layer, and ReLU activation function are regarded as a unit network layer. The SD-ResNet model has a total of L network layers, and the parameter θ of the SD-ResNet model is:

[0049] ,

[0050] The parameters of the lth layer network are: W l is the convolution kernel weight matrix of the lth layer; b l is the bias term of the lth layer; γ l is the scaling parameter of batch normalization of layer l; β l is the offset parameter of batch normalization of the lth layer, l <L。

[0051] The data enters the residual layer through the input layer and passes through the residual blocks in the residual layer. Each residual block decides whether to skip or not according to the random depth mechanism.

[0052] During training, in each forward propagation, each residual block will have a probability value to determine whether to skip the block. If the randomly generated number is less than the probability value, the residual block will be skipped and directly returned to the input without any convolution or activation operations. If it is not skipped, the convolution, normalization and other operations of the residual block are performed according to the normal process. The following is the specific training process:

[0053] The gradient of the loss function M with respect to the model parameters θ is calculated by backpropagation, and the parameters θ are updated according to these gradients using gradient descent; let the input of the lth layer be the output a of the l−1 layer (l−1) , the weight matrix is ​​W (l) , with a bias of b (l) ; Calculate linear combinations :

[0054] ,

[0055] Then the output a of the lth layer is obtained through the ReLU activation function (l) :

[0056] ,in represents the activation function,

[0057] The output of the last layer of the network is a (L) is the network prediction value ;

[0058] The loss function M is:

[0059] , N is the number of samples;

[0060] When calculating the hidden layer gradient forward layer by layer, for the lth layer, the error term δ(l) needs to be calculated:

[0061] ,

[0062] in, represents element-by-element multiplication, f′(z (l) ) is the activation function The derivative of .

[0063] Gradient calculation: The partial derivatives of the loss function M with respect to weights and biases are calculated as follows:

[0064] ,

[0065] Perform updates on each parameter using gradient descent:

[0066] ,

[0067] in, , for , The derivative of is the learning rate, which is set to 0.05 in this embodiment;

[0068] The data processed by the residual layer is the feature vector, which is input into the output layer, and the final prediction value is output through the output layer according to the target number of classification tasks. Finally, the post-processing module is used to restore the output of 16QAM data, and SD-ResNet learns the features of 16QAM data.

[0069] In general, this model uses the back-propagation algorithm to calculate the gradient of the loss function for each parameter and use these gradients to update the parameters, thereby minimizing the loss function and improving the performance of the model.

[0070] S4. Obtain equalized 16QAM data according to the equalized real data and the equalized imaginary data.

[0071] In the ResNet network of this solution, the residual idea is applied to the symbol recovery of the coherent optical communication system. ResNet is used to construct the IQ nonlinear mapping between the received symbols and the transmitted symbols, which solves the network degradation problem while maintaining a high learning ability. In addition, a random deep screening mechanism is added to the residual block to reduce the network training cost when facing a huge amount of data training, so as to achieve symbol balance decision. The mapping of the expected transmitted data and received data is denoted as H(x), and the residual relationship obtained by the residual block learning is F(x)=H(x)-x, that is, the relationship between the network input and output is H(x)=F(x)+x, and the input symbol matrix is ​​Matrix Tx i , the symbol matrix output through the optical fiber link is Matrix Rx i (i is the number of symbols in the matrix). Both matrices are complex matrices. i As the target to be restored, the data collected on the system ranges from -8dBm to -4dBm, and the data of these groups are unified as Matrix Rx i Class, using ResNet to convert the data change range of the corresponding points from Matrix Rx i To Matrix Tx i Residual learning.

[0072] In order to further reduce the complexity of network learning, a stochastic depth probability decision mechanism is introduced in the res-block. First, the real and imaginary parts of the 16QAM symbol stream are taken to form a one-dimensional (1D) vector as the input of the ResNet equalizer. ResNet will focus the limited network learning ability on this part of the strongly changing symbols, reducing the waste of training resources for weakly changing symbols.

[0073] Introducing a probabilistic screening mechanism in a deep learning model can improve the accuracy of the model in a variety of ways. Here is a common method: Stochastic Depth. Stochastic Depth is a regularization method that improves the training efficiency and performance of deep neural networks. Its basic idea is to randomly skip certain residual blocks during the training process so that the depth of the network is different for each training iteration, and the complete network is used in the test phase. This method can not only reduce the computational burden of the deep network, but also prevent overfitting, thereby improving the generalization ability of the model. The specific principle is as follows Figure 3 :

[0074] Design of residual blocks: During training, for each residual block, the computation of the block is skipped with a certain probability. If a residual block is skipped, the input is passed directly to the output without going through the computation of the block. Skip probability: During training, each residual block has a predefined skip probability p l Generally, residual blocks in deeper layers have higher skipping probability, while residual blocks in shallow layers have lower skipping probability.

[0075] For greater flexibility, the skip probability can be varied as the number of network layers increases. Usually, a higher survival probability is set for shallow residual blocks, and a lower survival probability is set for deep residual blocks. This approach can make shallow residual blocks more stable and deep residual blocks more random, thereby enhancing the regularization effect. For the mth residual block, the skip probability is:

[0076] ,

[0077] Where: p d is the skip probability of the dth residual block, p 0 is the initial skip probability, m is the index of the current residual block, and D is the total number of residual blocks; the skip probability decays linearly with the number of layers, ensuring that the residual blocks in shallow layers are more stable and the residual blocks in deep layers are more randomized.

[0078] In deep networks, gradient vanishing or gradient exploding problems may cause training difficulties. The random depth mechanism (SD) reduces the length of the forward and backward propagation paths by skipping some residual blocks, thereby improving the gradient flow, allowing the gradient to be more effectively propagated to shallow layers and improving training efficiency. The SD-ResNet proposed in this application dynamically adjusts the network depth so that the training process of each batch is slightly different, thereby introducing data augmentation effects during training, improving the robustness of the model, and preventing the model from relying on specific paths or specific parameters. In the test phase, all residual blocks are activated (that is, the random depth mechanism is not embedded, Figure 4Tra represents the original ResNet network without SD embedding), and the complete ResNet network is used for inference. d The network performance obtained by taking different values ​​is as follows Figure 3 As shown in the figure, it can be seen that when the survival probability of each layer is 0.7, the SD-resnet-block model has the best learning performance for IQ data, which is better than the ResNet network without the SD mechanism embedded in the same situation, and has the best convergence performance. It verifies that SD-ResNet has a shallower depth and guarantees good convergence performance.

[0079] During the inference process, if the random depth mechanism is also applied (such as skipping certain residual blocks with the same probability), the computational complexity will be reduced. Especially in time-sensitive sequences, this mechanism may bring significant performance improvements. The calculation complexity of SD-ResNet and ResNet network models mainly involves calculating the parameter quantity (number of parameters) and the amount of calculation (FLOPs, floating-point operations). Specifically, this involves the computational complexity of layers such as convolutional layers, batch normalization layers, and fully connected layers. Since the SD mechanism does not affect the parameter quantity, the parameter calculation formula remains unchanged. For a standard ResNet model, the parameter calculation formula is:

[0080]

[0081] Where P i is the number of parameters of the i-th layer, and L is the total number of layers.

[0082] The standard ResNet parameter calculation is:

[0083]

[0084] Among them: P represents the total number of parameters corresponding to each layer, Conv1 layer represents the convolutional layer, Layer i represents the i-th layer, and FC layer represents the fully connected layer.

[0085] The SD mechanism will reduce the actual amount of computation because some layers will be skipped in a certain forward propagation. It is necessary to introduce the survival probability when calculating the FLOPs of each layer. For FLOPs, it is necessary to introduce the survival probability 𝑝 in the calculation of the FLOPs of each layer. 𝑖 Assume that the FLOPs of the 𝑖th layer is 𝐹 𝑖 , then the FLOPs formula after introducing the survival probability is:

[0086]

[0087] where p iis the survival probability of the i-th layer.

[0088] For the SD-ResNet used in this application, the FLOPs calculation formula is:

[0089]

[0090] Where p=0.7 is the survival probability, F layeri represents the FLOPs of the i-th residual block.

[0091] After calculation, the Total Parameters and Total FLOPs of SD-ResNet are 2272064 and 7766340 respectively. They are denoted as m and n respectively. The complexity comparison between Traditional-ResNet and SD-ResNet is shown in Table 1:

[0092] Table 1. Complexity of Traditional-ResNet and SD-ResNet

[0093]

[0094] After verification, the Total Parameters of SD-ResNet are 70% lower than those of Traditional-ResNet, and its Total FLOPs are 180% lower than those of Traditional-ResNet. This shows that while ensuring convergence performance, the complexity of SD-ResNet is much lower than that of Traditional-ResNet, and the introduced SD mechanism has a significant improvement in the performance of the ResNet network.

[0095] In order to further evaluate the generalization ability of the ResNet model on unknown data. Figure 5 As shown, this application adopts the k-cross validation method, the basic idea of ​​which is to randomly divide the original data into k roughly equal subsets, and then take one of the subsets as the test set and the remaining k-1 subsets as the training set in turn, repeat this process k times, and finally take the mean of the k-times model evaluation results as the final evaluation index. The performance evaluation process and model accuracy calculation formula are as follows:

[0096]

[0097] Where D correctrepresents the number of samples predicted correctly, and D represents the total number of samples. Suppose there are N samples, which are divided into k roughly equal subsets, each containing approximately N / k samples. For the i-th validation (i=1,2,...,k): the i-th subset is used as the test set, denoted as D_test i, and the remaining k-1 subsets are combined as the training set, denoted as D_train i. The model is trained on D_train i to obtain the model Mi, and the model Mi1 is evaluated on D_test i to obtain the evaluation index accuracy Accuracy i. The final model evaluation result Accuracy is the mean of the k-times evaluation index: Accuracy = (Accuracy1 + Accuracy 2 + ... + Accuracy k) / k. Through k cross-validation, we can make full use of limited data, reduce the bias and variance of model evaluation, and obtain a more reliable estimate of the model's generalization ability. (The value of k is 10.)

[0098] This equalization scheme has been verified on a 25km coherent experimental platform. First, experiments were conducted on a coherent optical transmission system: at the transmitter, a tunable external cavity laser (ECL) with a linewidth of 100 kHz was used to transmit an optical signal with a wavelength of 1550 nm. After passing through a polarization beam splitter (PBS), the optical carrier was modulated by an IQ modulator (IQM). The IQM is driven by a four-channel arbitrary waveform generator (AWG) with an operating speed of 5 GSample / s, generating a 16QAM signal at 2.5 GBaud.

[0099] After passing through a polarization beam combiner (PBC), the two modulated signals are transmitted in a 20km standard single-mode fiber (SMF). First, an erbium-doped fiber amplifier (EDFA) is used to amplify the optical signal at the receiving end. Then, EDFA-2 is used as a noise source and enters the receiver through an optical coupler (OC) along with the signal light. A variable optical attenuator (VOA) is used to adjust the noise. After coherent reception by the local oscillator (LO), the four-channel PDM signal is digitized by a high-speed digital storage oscilloscope (DSO) with an acquisition rate of 50 GS / s. The acquired signal is processed by an offline digital signal processor (DSP).

[0100] The DSP part compares the traditional CMA algorithm and the SD-ResNet equalization scheme. In the original equalization algorithm, the least mean square-based multiple-input multiple-output (MIMO) algorithm is used for polarization demultiplexing, and the demodulated bit error rate (BER) is used as the performance indicator of the system. When using the SD-ResNet equalization scheme, the I, Q of the X channel and the I, Q of the Y channel are used to train the SD-ResNet neural network respectively. The trained model is used to perform phase recovery on 10 sets of validation set data, and the demodulated BER is also used as the performance indicator.

[0101] When the survival probability of the residual block is 0.7, the batch size ranges from 16 to 1024 and the training and testing are performed. Figure 6 The SD in stands for SD-ResNet, and the number indicates the size of the batch size. The results show that as the batch size increases, the loss value of the network decreases faster. This is because a larger batch size can smooth the loss curve to a certain extent, provide a smaller gradient estimation variance, and thus make the gradient estimation more accurate, the network parameter optimization path more stable, and reduce the impact of noise on training. In addition, as the number of epochs increases, the loss function value gradually decreases, thanks to the SD-ResNet model gradually updating the parameters through back-propagation and gradient descent optimization algorithms. In each epoch, the model adjusts the parameters by calculating the gradient of the loss function to gradually reduce the loss value. However, from Figure 6 As can be seen from the figure, when the batch size increases to 1024, the convergence performance of the network decreases. Too large a batch size causes greater changes in the gradient direction, making the training process more prone to oscillations and jumps, resulting in unstable learning. Compared with SD-ResNet with other batch sizes, SD512 has the best convergence performance, and the loss function value is always at the lowest state.

[0102] In terms of the generalization ability of the model, the k-cross validation method was used to evaluate the performance of four SD-ResNet models (image-Rx, image-Ry, Real-Rx, Real-Ry in the figure). Figure 7 As shown, image-Rx and image-Ry represent the models for training imaginary data, while Real-Rx and Real-Ry represent the models for training real data. First, divide the dataset into 10 subsets of equal size (k is 10). Select a training set and a validation set. In the i-th iteration, select the i-th subset as the validation set and the remaining 9 subsets as the training sets. Train the model on the selected training set, evaluate the model performance on the validation set, and record the accuracy of the model. After completing 10 iterations, calculate the average of the accuracy recorded in all iterations. These average performance indicators can more reliably reflect the actual performance of the model. As Figure 7 As shown in the figure, the accuracy of the four SD-ResNets on different validation sets is higher than 98%, and the average accuracy of the model corresponding to Real-Ry is 99.66%. This proves that the ResNet model with SD mechanism has good generalization ability in learning coherent data features.

[0103] The above evaluation of the structural performance of the model. In order to test the effectiveness of the model in learning nonlinear features of IQ data, the following Figure 8 The IQ data transmitted through the 25km coherent system was processed, the real data and the imaginary data were separated, and different SD-ResNet networks were trained respectively. Four SD-ResNet equalization models corresponding to the real X, Y data and the imaginary X, Y data were obtained respectively. The feature extraction and recovery of the data with the input fiber power ranging from -8dbm to -4dBm were performed respectively. There were 10 groups of validation data corresponding to each input fiber power, and the FEC threshold was set to 3.8×10 -3 , the BER performance of the system is as follows Figure 8 , where Tra-CAM represents the traditional CMA equalization solution, and SD-ResNet represents the solution proposed in this patent. Average represents the BER average under the same power.

[0104] The results of the experiment are as follows Figure 8 As shown in the figure, in terms of IQ data feature learning effect, with FEC=3.8×10 -3 As the BER judgment standard, the received optical power corresponding to the Tra-CMA scheme is -6.14dBm, and the received optical power corresponding to the SD-ResNet is -6.8dBm. Therefore, the receiver sensitivity gain provided by the SD-ResNet equalization is 0.66dB compared with the data after the traditional CMA algorithm equalization. This shows that the proposed SD-ResNet has a significant effect on the IQ data recovery of the 25km coherent system and can effectively improve the system's bit error rate performance.

[0105] Embodiment 2: Based on Embodiment 1, this embodiment provides a coherent optical nonlinear IQ impairment equalization device, including a processor and a storage medium;

[0106] The storage medium is used to store instructions;

[0107] The processor is used to operate according to the instruction to execute the method according to embodiment 1.

[0108] Embodiment 3: Based on Embodiment 1, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described in Embodiment 1 is implemented.

[0109] Embodiment 4: Based on Embodiment 1, this embodiment provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described in Embodiment 1 when executing the computer program.

[0110] Embodiment 5: Based on Embodiment 1, this embodiment provides a computer program product, including a computer program, which implements the method described in Embodiment 1 when executed by a processor.

[0111] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0112] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of the processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0115] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for equalizing coherent optical nonlinear IQ impairments, characterized in that: include: S1. Obtain 16QAM data to be equalized; S2, preprocessing the 16QAM data to obtain preprocessed real data and imaginary data; S3, inputting the preprocessed real data into the pre-trained first SD-ResNet model to obtain the balanced real data; The preprocessed imaginary part data is input into the pre-trained second SD-ResNet model to obtain the balanced imaginary part data; S4, obtaining equalized 16QAM data according to the equalized real part data and the equalized imaginary part data; The method for obtaining the pre-trained first SD-ResNet model and the second SD-ResNet model includes: The network structures of the first SD-ResNet model and the second SD-ResNet model are the same, and both include: an input layer, four residual layers and an output layer; the input layer includes a 1-dimensional convolution layer, a batch normalization layer and a ReLU activation function; each residual layer includes two residual blocks, and each residual block includes a 1-dimensional convolution layer, a batch normalization layer and a ReLU activation function; the output layer includes a global average pooling layer, a flattening layer and a fully connected layer; The first SD-ResNet model is trained using a labeled real data sample set, and the second SD-ResNet model is trained using a labeled imaginary data sample set until the preset conditions are met to obtain the trained first SD-ResNet model and the second SD-ResNet model; during the training process, for each residual block, a decision is made whether to skip it based on a random depth mechanism. If the randomly generated number is less than the survival probability of the residual block, the calculation of the residual block is skipped and the input is directly returned; if the randomly generated number is greater than or equal to the survival probability of the residual block, the residual block is calculated.

2. The method according to claim 1, characterized in that The survival probability of each residual block is 0.

7.

3. The method according to claim 1, characterized in that The data processing process of the first SD-ResNet model and the second SD-ResNet model includes: The preprocessed real data and imaginary data are downsampled by the input layer and then enter the residual layer. In the residual layer, the random depth mechanism is used to decide whether to skip the calculation of the residual block. The feature vector is obtained by processing the residual layer. The feature vector enters the output layer and is processed by the global average pooling layer, the flattening layer, and the fully connected layer to obtain the balanced real data and imaginary data.

4. The method according to claim 1, characterized in that: The training process of the first SD-ResNet model and the second SD-ResNet model includes: The connected convolutional layer, batch normalization layer and ReLU activation function are regarded as a unit network layer. The SD-ResNet model has a total of L network layers. The parameter θ of the SD-ResNet model is: ， The parameters of the lth layer network are: W l is the convolution kernel weight matrix of the lth layer; b l is the bias term of the lth layer; γ l is the scaling parameter of batch normalization of layer l; β l is the offset parameter of batch normalization of the lth layer, l <L; Let the input of layer l be the output a of layer l−1 (l−1) , the weight matrix is ​​W (l) , with a bias of b (l) ; Calculate linear combinations : ， Then the output a of the lth layer is obtained through the ReLU activation function (l) : ,in represents the activation function, The output of the last layer of the network is a (L) is the network prediction value .

5. The method according to claim 4, characterized in that The training process of the first SD-ResNet model and the second SD-ResNet model further includes: The gradient of the loss function M with respect to the model parameter θ is calculated by back propagation, and the parameter θ is updated according to these gradients using gradient descent, thereby minimizing the loss function M to obtain the trained first SD-ResNet model and the second SD-ResNet model; the loss function M is: ， Where N is the number of samples; x i is the ith real or imaginary data input sample, y i is the true output of the i-th input sample, and f(xi;θ) is the predicted output of the SD-ResNet model; When calculating the hidden layer gradient layer by layer, for the lth layer, calculate the error term δ (l) : ， Among them, δ (l) ,δ (l+1) represents the error term of the lth and l+1th layers, represents element-by-element multiplication, f′(z (l) ) is the activation function The derivative of Gradient calculation: The partial derivatives of the loss function M with respect to weights and biases are calculated as follows: ， Perform updates on each parameter using gradient descent: ， in, , for , The derivative of is the learning rate of the network.

6. The method according to claim 1, characterized in that S2, preprocessing the 16QAM data to obtain real data and imaginary data after preprocessing; including: Performing Z-standardization processing on the real data and the imaginary data of the 16QAM data respectively to obtain standardized real data and imaginary data; The standardized real data and imaginary data are respectively dimensionally expanded to obtain preprocessed real data and imaginary data.

7. A coherent optical nonlinear IQ impairment equalization device, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

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

Images