An intelligent equalization algorithm for optical communication based on physical adjustment

Abstract

The application discloses a kind of optical communication intelligent equalization algorithm based on physical adjustment, including steps S1: initialization neural network;Step S2: carry out forward propagation;Step S3: carry out loss function calculation;Step S4: carry out gradient calculation;Step S5: carry out back propagation;Step S6: carry out neural network variable value analysis;Step S7: reinitialize neural network by estimated physical parameter.The application discloses a kind of optical communication intelligent equalization algorithm based on physical adjustment, traditional static DSP is parameterized as trainable neural network, and then the intelligent update of neural network variable is carried out using optimization algorithm.At the same time, combined with physical knowledge is applied to neural network and neural network information feedback physics, the effect of mutual promotion of both is achieved, so that high-precision parameter estimation and signal impairment compensation can be carried out on optical communication system with unknown parameters.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical communication, and particularly relates to an optical communication intelligent equalization algorithm based on physical adjustment. BACKGROUND

[0002] Optical communication technology is widely used and greatly changes the mode of data transmission. Taking optical fiber communication technology as an example, it has the advantages of extremely low loss, excellent bandwidth capacity and long-distance transmission capability. However, the transmission of optical signals in a communication channel will produce damage, thereby deteriorating the transmission quality of the signals. Therefore, the optical communication system needs to use a digital signal processing (DSP) algorithm to implement the equalization of the signals to recover the damage produced in the transmission process, thereby improving the transmission quality.

[0003] Digital back propagation (DBP) as a traditional DSP method has high accuracy, but the parameters do not change after initialization. This static equalization algorithm is sensitive to the parameter fluctuations of the system and is also difficult to perform channel equalization on unknown systems. Neural networks have the ability to learn from training data, so they have great potential in adaptive damage compensation of unknown systems. For example, Freire et al. published a paper entitled "Computational Complexity Optimization of Neural Network-Based Equalizers in Digital Signal Processing: A Comprehensive Approach" in the Journal of Lightwave Technology in 2024, which discusses the effects of using various neural network architectures to perform channel equalization on coherent optical communication systems. The analysis shows that in order to achieve good equalization effect, the neural network algorithm needs to be trained with high complexity, which greatly increases the computing resources consumed by the algorithm and the time required for the algorithm to run. Similarly, et al. published a paper entitled "Physics-Based Deep Learning for Fiber-Optic Communication Systems" in the IEEE Journal on Selected Areas in Communications in 2021, which proposed a learning DBP (LDBP) method. By parameterizing the DBP method as a neural network, an excellent initialization network architecture is constructed, thereby improving the convergence speed of the network and achieving better equalization effect. However, this method still needs to know the relevant optical communication system parameters in advance when initializing the variables.

[0004] In summary, the existing algorithm is better in the case of accurate parameters of optical communication system. However, the parameters of optical communication system are not necessarily determined in advance. Therefore, it is necessary to provide an intelligent equalization algorithm which can still perform well in unknown optical communication system. SUMMARY

[0005] The main purpose of the present application is to provide an optical communication intelligent equalization algorithm based on physical regulation, which parameterizes the traditional static DSP into a trainable neural network, and then uses an optimization algorithm to intelligently update the neural network variables. At the same time, the physical knowledge is applied to the neural network and the neural network information feedback is physical, so as to achieve the mutual promotion effect of the two, so as to perform high-precision parameter estimation and signal damage compensation on the optical communication system with unknown parameters.

[0006] To achieve the above purpose, the present application provides an optical communication intelligent equalization algorithm based on physical regulation, comprising the following steps:

[0007] Step S1: initializing the neural network, initializing the structure of the neural network as a trainable variable;

[0008] Step S2: forward propagation, equalizing the damaged signal according to the current neural network variable value, the input of which is the damaged signal transmitted by the optical communication system, and the output of which is the equalized signal, the target output of which is consistent with the original signal before the optical communication transmission;

[0009] Step S3: loss function calculation, comparing the equalized signal obtained by forward propagation with the real original signal, and calculating the loss value in the loss function;

[0010] Step S4: gradient calculation, calculating the gradient of the variable in the neural network to the loss value through the loss value calculated by the loss function;

[0011] Step S5: back propagation, using the calculated gradient to optimize the variable value in the neural network through the optimization algorithm;

[0012] Step S6: neural network variable value analysis, extracting the physical law in the optimized neural network variable by analyzing the optimized neural network variable, and realizing the estimation of the physical parameters;

[0013] Step S7: reinitializing the neural network through the estimated physical parameters.

[0014] As a further preferred technical solution of the above technical solution, for step S1:

[0015] Step S1.1: linear operator in the DSP algorithm is initialized as a variable linear weight in the neural network, denoted as weight W, so as to apply physical knowledge to the neural network;

[0016] Step S1.2: the static nonlinear operator in the DSP algorithm is initialized as a variable nonlinear activation function in the neural network, denoted as function f, so as to apply physical knowledge to the neural network;

[0017] Step S1.3: in the neural network thus initialized, one or more linear weight modules and one or more nonlinear activation function modules are combined to form a layer of neural network; the function of each layer of neural network is written as x = f(Wx0), wherein x is the output of the layer of neural network, and x0 is the input of the layer of neural network.

[0018] As a further preferred technical solution of the above technical solution, for step S2, the forward propagation is used for calculation of the loss function in the training process; the output of the forward propagation in the inference process represents the predicted real original signal, and compensation of the signal impairment is realized.

[0019] As a further preferred technical solution of the above technical solution, for step S3, the loss value calculated by selection of the loss function reflects the result of the forward propagation; the loss function is denoted as f L , which is selected as a mean square error loss function; L = ||y-y t || 2 / N, y t is a real original signal vector, becomes an impaired signal y0 after passing through a communication system, and becomes y after passing through the forward propagation, which is an equalized signal vector obtained by the forward propagation, |||| represents a modulus operation on a vector, N is the length of the two signals, and L is the calculated loss value.

[0020] As a further preferred technical solution of the above technical solution, for step S4, the gradient calculation step calculates the gradient of the variable value of the trainable neural network initialized after the neural network with respect to the loss value, for a single layer of neural network, the gradient needs to be calculated according to the derivative of the loss value L with respect to the weight W, and is denoted as:

[0021] As a further preferred technical solution of the above technical solution, for step S5, the variable optimization of the back propagation process is realized by using an Adam optimizer.

[0022] As a further preferred technical solution of the above technical solution, for step S6, the neural network variable value analysis process reads out the physical parameters corresponding to the neural network variables in the original DSP algorithm, realizes the estimation of the physical parameters, and realizes the information feedback from the neural network to the physics; when the neural network is initialized, if the variables of a single neural network correspond to a single physical parameter, the variables after the back propagation optimization are directly read out as the physical parameters; if the variables of multiple neural networks correspond to a single physical parameter, the multiple neural network variables are averaged and then read out as the physical parameters.

[0023] As a further preferred technical solution of the above technical solution, for step S7, the re-initialization process makes the physical adjustment-based optical communication intelligent equalization algorithm perform multiple cycles, sets an upper limit of the cycle of the training process to control the number of cycles, and re-initializes the physical parameters used to be closer to the true value than the last initialization. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is the algorithm flowchart of the present application.

[0025] Figure 2 is the schematic process diagram of initializing the neural network of the present application.

[0026] Figure 3 is the training process diagram of the present application applied to an optical communication system.

[0027] Figure 4 is the inference process of the present application applied to an optical communication system.

[0028] Figure 5 is the constellation diagram and the comparison diagram of the conventional method when the present application is applied to an optical communication system.

[0029] Figure 6 is the bit error rate and the comparison diagram of the conventional method when the present application is applied to an optical communication system.

[0030] Figure 7 is the physical parameter estimation diagram of the present application applied to an optical communication system. DETAILED DESCRIPTION

[0031] The following description is provided to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only examples, and other obvious modifications can be made by those skilled in the art. The basic principles of the present application defined in the following description can be applied to other embodiments, modifications, improvements, equivalents and other technical solutions without departing from the spirit and scope of the present application.

[0032] In preferred embodiments of the present application, the skilled person will note that the communication system and the like to which the present application relates can be considered as prior art.

[0033] Preferred embodiments.

[0034] As Figures 1-7 shown, the present application discloses a physical regulation-based optical communication intelligent equalization algorithm, comprising the following steps: the application of the algorithm is divided into two processes of training and reasoning. Among them, in the training process, all the following steps need to be performed to get the best effect; in the reasoning process, only the forward propagation step needs to be performed, so the remaining processes are represented by dashed lines in Figure 1 .

[0035] Step S1: initialize the neural network, (refer to the conventional DSP algorithm) initialize the structure of the neural network as a trainable variable (initialize the structure and variable value of the network as the structure and value of the conventional DSP);

[0036] Step S2: forward propagation, according to the current neural network variable value, the equalization of the damaged signal is carried out, the input is the damaged signal transmitted through the optical communication system, and the output is the equalized signal, the equalized signal of the target output is consistent with the original signal before the optical communication transmission;

[0037] Step S3: loss function calculation, compare the equalized signal obtained by forward propagation with the real original signal, and calculate the loss value in the loss function;

[0038] Step S4: gradient calculation, calculate the gradient of the variable in the neural network to the loss value through the loss value calculated by the loss function;

[0039] Step S5: back propagation, using the calculated gradient, the variable value in the neural network is optimized through the optimization algorithm;

[0040] Step S6: neural network variable value analysis, by analyzing the optimized neural network variable, the physical law is extracted, and the estimation of the physical parameter is realized;

[0041] Step S7: reinitialize the neural network through the estimated physical parameter (again refer to the conventional DSP algorithm).

[0042] Specifically, for step S1:

[0043] Step S1.1: (the schematic diagram of initializing the neural network is shown in Figure 2Step S1.1: initialize the static linear operator in the DSP algorithm to the variable linear weight in the neural network, denoted as weight W (for example, the dispersion operator in the DBP method can be initialized to the linear weight in the neural network), so that the physical knowledge is applied to the neural network;

[0044] Step S1.2: initialize the static nonlinear operator in the DSP algorithm to the variable nonlinear activation function in the neural network, denoted as function f (for example, the nonlinear operator in the DBP method can be initialized to the nonlinear activation function in the neural network), so that the physical knowledge is applied to the neural network;

[0045] Step S1.3: in the neural network thus initialized, the combination of one or more linear weight modules and one or more nonlinear activation function modules after them is a layer of neural network (for example, in the classic DBP algorithm, the data processing is alternately through the linear dispersion operator and the nonlinear operator, and after they are initialized to the neural network, the combination of one dispersion operator and one nonlinear operator is a layer of neural network. The alternately passing through two operators in the original DBP is changed to passing through multiple layers of neural network one by one); the effect of each layer of neural network is written as x = f (Wx0), where x is the output of the current layer of neural network, and x0 is the input of the current layer of neural network.

[0046] wherein the value of the DSP needs the physical parameter value of the optical communication system, such as the dispersion and nonlinear coefficient of the optical fiber; therefore, for an unknown optical communication system, a set of default physical parameters can be used to calculate the value of the DSP, and then the set of values is used to initialize the variables of the neural network for the first time; in Figure 2 , the static DSP represents the traditional channel equalization method, and the internal modules thereof are represented by double-layer solid lines; the variable neural network represents the proposed optical communication intelligent equalization algorithm based on physical adjustment, and the internal modules thereof are represented by three-layer solid lines; the action modes of both are to take the damaged signal as the input and take the equalized signal as the output.

[0047] More specifically, for step S2, the forward propagation is used for the calculation of the loss function in the training process (the output result thereof may not be consistent with the true original signal, and the cooperation of other processes in the optical communication intelligent equalization algorithm based on physical adjustment is needed); the output of the forward propagation in the inference process represents the predicted true original signal, and the compensation of the signal damage is realized (other processes in the optical communication intelligent equalization algorithm based on physical adjustment are no longer needed).

[0048] Further, for step S3, the loss value calculated by the selection of the loss function reflects the result of the forward propagation; the loss function is denoted as f L , and the selection is the mean square error loss function; L = ||y-y t || 2 / B, y tFor the real original signal vector (corresponding to the original signal in Figure 3 , which becomes the impaired signal after going through the communication system, denoted as y0, and becomes y after going through the forward propagation, which is the equalized signal vector after forward propagation (corresponding to the neural network output in Figure 3 , and represents the modulo operation on the vector, N is the length of the two signals, and L is the calculated loss value (the calculation of y is obtained by the forward propagation of y0 through the multi-layer neural network, so the multi-layer linear weight and the activation function are included in L, and the gradient of W and f in different layers can be analyzed through L).

[0049] Further, for step S4, the gradient calculation step calculates the gradient of the variable values of the neural network initialized for training with respect to the loss value. For a single-layer neural network (as an example), the gradient needs to be calculated according to the derivative of the loss value L with respect to the weight W, denoted as:

[0050] Preferably, for step S5, the variable optimization of the backpropagation process is implemented using the Adam optimizer (for example, the basic optimization formula is: where W u is the optimized weight, R l is the optimized learning rate, is the calculated gradient of L with respect to the weight W).

[0051] Preferably, for step S6, the neural network variable value analysis process reads out the neural network variables corresponding to the physical parameters (such as the dispersion and nonlinear coefficient of the optical fiber) in the original DSP algorithm, realizes the estimation of the physical parameters, and realizes the information feedback from the neural network to the physical; when the neural network is initialized, if a single neural network variable corresponds to a single physical parameter, the variable after optimization by backpropagation is directly read out as the physical parameter; if multiple neural network variables correspond to a single physical parameter, the multiple neural network variables are averaged and then read out as the physical parameter (since the optimizer generally moves the loss value in the direction of reduction, the optimized variable corresponding to the physical parameter in the original DSP will be closer to the true value than before optimization).

[0052] Preferably, for step S7, the re-initialization process makes the physical adjustment-based optical communication intelligent equalization algorithm perform multiple cycles, sets an upper limit for the training process to control the number of cycles, and re-initializes the physical parameters used to be closer to the true value than the last initialization (therefore, this will provide a better initialization for the physical adjustment-based optical communication intelligent equalization algorithm, thereby promoting better results in the next cycle).

[0053] In this process, physical regulation is reflected in the combination of knowledge application from physics to neural network and information feedback from neural network to physics, corresponding to the steps of "initializing neural network" and "neural network variable value analysis". Through the combination of the two, the neural network after each cycle can obtain better initialization and thus better performance. Traditional neural network optimization algorithms often focus on optimizing the value of the loss function, ignoring the internal physical laws, leading to overfitting. The proposed optical communication intelligent equalization algorithm based on physical regulation can alleviate the adverse effects of overfitting due to the presence of physical regulation.

[0054] The training process of the optical communication intelligent equalization algorithm based on physical regulation applied to an optical communication system is shown in Figure 3 To train, a known original signal needs to be input into the optical communication system to obtain a damaged signal after transmission; the damaged signal is input into the optical communication intelligent equalization algorithm based on physical regulation which has been initialized for the first time, and then forward propagation is performed to obtain the output; the neural network output is compared with the known original signal to calculate the loss value, and the corresponding gradient and back propagation are used to optimize the variable value of the neural network; the optimized variable value is analyzed to estimate the true physical parameters, and then the neural network is initialized again according to the physical parameters, and the cycle of forward propagation, loss function calculation, gradient calculation, back propagation, variable value analysis, and re-initialization of the neural network is continued until the upper limit of the set cycle is reached.

[0055] The inference process of the optical communication intelligent equalization algorithm based on physical regulation applied to an optical communication system is shown in Figure 4 The unknown original signal generates a damaged signal through the optical communication system, and at this time the optical communication intelligent equalization algorithm based on physical regulation has been trained, so only a single forward propagation of the damaged signal is needed to obtain the equalized signal.

[0056] When the optical communication intelligent equalization algorithm based on physical regulation is applied to an optical fiber communication system, the optical fiber communication system parameters are set to 10x100km transmission of standard single mode fiber (SMF), quadrature amplitude modulation (QAM) format, fiber input power of 0dBm, optical signal-to-noise ratio of 15dB, transmission rate of 32Gbaud, and symmetrically distributed Fourier method with step size of 1km is used for simulation. In order to simulate the situation of equalization under unknown fiber parameters, the initial parameters of DBP, LDBP and physical regulation neural network algorithm are all set to add perturbation to the parameters of SMF. The greater the perturbation, the worse the initialization of the algorithm and the more serious the mismatch with the real situation. For neural network-based equalization algorithms (LDBP and physical regulation neural network), the number of forward and backward propagations is uniformly set to 650 times to ensure fair performance comparison.

[0057] The constellation diagram obtained by the physical adjustment-based optical communication intelligent equalization algorithm is shown in Figure 5 The constellation diagram obtained by the LDBP is Figure 5 The left corresponds to the bit error rate of 9.11*10 -2 The constellation diagram obtained by the physical adjustment-based optical communication intelligent equalization algorithm is Figure 5 The right corresponds to the bit error rate of 4.39*10 -3 It can be found that the equalization effect of the physical adjustment neural network is obviously better than that of the traditional LDBP method.

[0058] The bit error rate curve obtained by the physical adjustment-based optical communication intelligent equalization algorithm is shown in Figure 6 Taking the equalization of the 16QAM signal as an example, the bit error rates obtained by the DBP, the LDBP and the physical adjustment neural network method are counted, and the values change with the mismatch of the β2 (dispersion) parameter at the initialization. It can be found that the equalization ability of the static DBP method for the unknown system is the worst, the LDBP method has certain adaptability due to the characteristics of the neural network, and the bit error rate of the proposed physical adjustment neural network equalization method is the lowest, and the effect is the best.

[0059] The physical parameter estimation of the physical adjustment-based optical communication intelligent equalization algorithm is shown in Figure 7 With the increase of the estimation times, the physical parameters (the blue solid line corresponding to Estimation) extracted by the algorithm gradually converge from the parameters at the initialization to the real parameters (the red dotted line corresponding to Ground-truth), which reflects the ability of the algorithm for the physical parameter estimation.

[0060] It is worth mentioning that the technical features of the communication system and the like involved in the present patent application should be regarded as the prior art, the specific structure, working principle and possible control mode and spatial arrangement mode of these technical features can be selected by using the conventional selection in the field, and should not be regarded as the invention point of the present patent, and the present patent will not be further specifically expanded and described.

[0061] For those skilled in the art, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced, any modification, equivalent replacement, improvement and the like within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A physical regulation based intelligent equalization algorithm for optical communication, characterized in that, The method comprises the following steps: Step S1: initializing the neural network, initializing the structure of the neural network as a trainable variable; Step S2: performing forward propagation, equalizing the impaired signal according to the current neural network variable value, the input of which is the impaired signal transmitted through the optical communication system, and the output of which is the equalized signal, the equalized signal of the target output being consistent with the original signal before optical communication transmission; Step S3: performing loss function calculation, comparing the equalized signal obtained by forward propagation with the real original signal, and calculating the loss value in the loss function; Step S4: performing gradient calculation, calculating the gradient of the variable in the neural network with respect to the loss value calculated by the loss function; Step S5: performing back propagation, using the calculated gradient to optimize the variable value in the neural network through an optimization algorithm; Step S6: performing neural network variable value analysis, extracting the physical law from the optimized neural network variable to estimate the physical parameter; Step S7: reinitializing the neural network based on the estimated physical parameter.

2. The physical regulation based optical communication intelligent equalization algorithm according to claim 1, wherein, For step S1: Step S1.1: initializing the static linear operator in the DSP algorithm as a variable linear weight in the neural network, denoted as weight W, so as to apply physical knowledge to the neural network; Step S1.2: initializing the static nonlinear operator in the DSP algorithm as a variable nonlinear activation function in the neural network, denoted as function f, so as to apply physical knowledge to the neural network; Step S1.3: in the neural network initialized in this way, one or more linear weight modules and one or more nonlinear activation function modules connected in sequence form a layer of neural network; the function of each layer of neural network is written as x=f(Wx0), wherein x is the output of the neural network, and x0 is the input of the neural network.

3. The physical regulation based optical communication intelligent equalization algorithm according to claim 2, wherein, For step S2, forward propagation is used for loss function calculation in the training process; the output of forward propagation in the inference process represents the predicted real original signal, realizing compensation of signal impairment.

4. The physical regulation based optical communication intelligent equalization algorithm according to claim 3, characterized in that, For step S3, the loss value calculated by the selection of the loss function reflects the forward propagation result; let the loss function be f L , which is selected as the mean square error loss function; L = || y - y t || 2 / N, y t is the real original signal vector, becomes the damaged signal after passing through the communication system, denoted as y0, and becomes y after passing through the forward propagation, which is the equalized signal vector obtained by the forward propagation, || || represents the modulus value operation on the vector, N is the length of the two signals, and L is the calculated loss value.

5. The physical regulation based optical communication intelligent equalization algorithm according to claim 4, characterized in that, For step S4, the gradient calculation step calculates the gradient of the values of the variables that are trainable after the initialization of the neural network with respect to the loss value. For a single-layer neural network, the gradient needs to be calculated based on the derivative of the loss value L with respect to the weight W, denoted as:

6. The physical regulation based optical communication intelligent equalization algorithm according to claim 5, wherein, For step S5, the variable optimization in the back propagation process is realized by using the Adam optimizer.

7. The physical regulation based optical communication intelligent equalization algorithm according to claim 6, wherein, For step S6, the neural network variable value analysis process reads the physical parameter corresponding to the original DSP algorithm from the neural network variable, realizes estimation of the physical parameter, and feeds back information from the neural network to the physical parameter; in the neural network initialization, if a single neural network variable corresponds to a single physical parameter, the variable optimized by back propagation is directly read out as the physical parameter; if multiple neural network variables correspond to a single physical parameter, the multiple neural network variables are averaged and then read out as the physical parameter.

8. The physical regulation based optical communication intelligent equalization algorithm according to claim 7, wherein, For step S7, the reinitialization process enables the optical communication intelligent equalization algorithm based on physical adjustment to be cycled multiple times, sets an upper limit for the training process to control the number of cycles, and uses the physical parameter closer to the real value than the previous initialization in the reinitialization process.

Images