Optical communication intelligent equalization algorithm based on physical adjustment

Abstract

The invention discloses an optical communication intelligent equalization algorithm based on physical adjustment. The algorithm comprises the following steps: S1, initializing a neural network; s2, carrying out forward propagation; s3, carrying out loss function calculation; s4, gradient calculation is carried out; s5, carrying out back propagation; s6, performing neural network variable value analysis; and S7, re-initializing the neural network through the estimated physical parameters. According to the optical communication intelligent equalization algorithm based on physical adjustment disclosed by the invention, a traditional static DSP is parameterized into a trainable neural network, and then intelligent updating of neural network variables is carried out by using an optimization algorithm. Meanwhile, physical knowledge is applied to the neural network and neural network information feedback physics, the mutual promotion effect of the neural network and the neural network information feedback physics is achieved, and therefore high-precision parameter estimation and signal damage compensation can be conducted on an optical communication system with unknown parameters.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical communication, and in particular relates to an optical communication intelligent balancing algorithm based on physical regulation. Background Art

[0002] Optical communication technology is widely used and has greatly changed the mode of data transmission. Taking optical fiber communication technology as an example, its advantages are extremely low loss, excellent bandwidth capacity and long-distance transmission capability. However, optical signals will be damaged during the transmission process of the communication channel, thereby deteriorating the transmission quality of the signal. Therefore, optical communication systems need to use digital signal processing (DSP) algorithms to implement signal equalization to restore the damage caused during the transmission process, thereby improving the transmission quality.

[0003] Digital back propagation (DBP), as a traditional DSP method, has high accuracy, but its parameters no longer change after initialization. This static equalization algorithm is sensitive to the parameter fluctuations of the system, and it is difficult to perform channel equalization on unknown systems. Neural networks have the ability to learn based on training data, so they have great potential in adaptive damage compensation for unknown systems. For example, Freire et al. published a paper titled "Computational Complexity Optimization of Neural Network-Based Equalizers in Digital Signal Processing: A Comprehensive Approach" in the journal Journal of Lightwave Technology in 2024, discussing the effects of using a variety of neural network architectures to perform channel equalization on coherent optical communication systems. The analysis shows that in order to achieve better equalization effects, the neural network algorithm needs to be trained with higher 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 titled "Physics-Based Deep Learning for Fiber-Optic Communication Systems" in the IEEE Journal on Selected Areas in Communications in 2021. They proposed a learning-based DBP (LDBP) method, which parameterized the DBP method into a neural network and constructed an excellent initialization network architecture, thereby improving the network's convergence speed and achieving better balancing effects. However, this method still requires the prior knowledge of the relevant optical communication system parameters when initializing the variables.

[0004] In summary, the existing algorithms work better when the parameters of the optical communication system are precisely known. However, the parameters of the optical communication system may not be predetermined. Therefore, it is necessary to provide an intelligent equalization algorithm that can still perform well in an unknown optical communication system. Summary of the invention

[0005] The main purpose of the present invention is to provide an optical communication intelligent equalization algorithm based on physical regulation, which transforms the traditional static DSP parameters 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 physics to achieve a mutual promotion effect between the two, so that high-precision parameter estimation and signal damage compensation can be performed on the optical communication system with unknown parameters.

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

[0007] Step S1: Initialize the neural network and initialize the structure of the neural network as trainable variables;

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

[0009] Step S3: Calculate the loss function, compare the equalized signal obtained by forward propagation with the real original signal, and calculate the loss value thereof through the loss function;

[0010] Step S4: perform gradient calculation, calculate the gradient of the variable in the neural network with respect to the loss value obtained by calculating the loss value through the loss function;

[0011] Step S5: Perform back propagation, and use the calculated gradient to optimize the variable values ​​in the neural network through an optimization algorithm;

[0012] Step S6: Analyze the neural network variable values, extract the physical laws from the optimized neural network variables, and estimate the physical parameters.

[0013] Step S7: Reinitialize the neural network using the estimated physical parameters.

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

[0015] Step S1.1: Initialize the static linear operator in the DSP algorithm to the variable linear weight in the neural network, denoted as weight W, so as to achieve the application of physical knowledge to the neural network;

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

[0017] Step S1.3: In the neural network initialized in this way, the combination of one or more linear weight modules followed by one or more nonlinear activation function modules forms a layer of neural network; the function of each layer of neural network is written as x=f(Wx0), where x is the output of this layer of neural network, and x0 is the input of this layer of neural network.

[0018] As a further preferred technical solution of the above technical solution, for step S2, forward propagation is used to calculate the loss function during the training process; the output of forward propagation in the reasoning process represents the predicted true original signal, thereby realizing compensation for signal damage.

[0019] As a further preferred technical solution of the above technical solution, for step S3, the loss value calculated by selecting the loss function reflects the forward propagation result; the loss function is denoted as f L , which is chosen as the mean square error loss function; L = ||yy t || 2 / N,y t is the real original signal vector, which becomes a damaged signal after passing through the communication system, denoted as y0, and then becomes y after forward propagation. is the equalized signal vector obtained by forward propagation, |||| represents the modulus operation on the vector, N is the length of the two signal segments, 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 uses the loss value to calculate the gradient of the trainable variable value after the neural network is initialized with respect to the loss value. For a single-layer neural network, its gradient needs to be calculated based on the derivative of the loss value L with respect to the weight W, which is expressed 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 implemented using the 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 the physical parameters in the original DSP algorithm corresponding to the neural network variables, realizes the estimation of the physical parameters, and feedbacks the information from the neural network to the physics; when the neural network is initialized, if the variable of a single neural network corresponds to a single physical parameter, the variable optimized by back propagation is directly read as the physical parameter; if the variables of multiple neural networks correspond to a single physical parameter, the multiple neural network variables are averaged and then read as the physical parameter.

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

[0024] Figure 1 It is a schematic diagram of the algorithm flow of the present invention.

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

[0026] Figure 3 It is a diagram of the training process when the present invention is applied to an optical communication system.

[0027] Figure 4 This is the reasoning process of the present invention when applied to an optical communication system.

[0028] Figure 5 It is a comparison diagram of the constellation diagram when the present invention is applied to an optical communication system and a traditional method.

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

[0030] Figure 7 It is a physical parameter estimation diagram when the present invention is applied to an optical communication system. DETAILED DESCRIPTION

[0031] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0032] In the preferred embodiments of the present invention, those skilled in the art should note that the communication system and the like involved in the present invention may be regarded as prior art.

[0033] Preferred embodiments.

[0034] like Figure 1-7 As shown, the present invention discloses an optical communication intelligent equalization algorithm based on physical regulation, including the following steps: the application of the algorithm is divided into two processes: training and reasoning. In the training process, all the following steps need to be performed to obtain the best effect; in the reasoning process, only the forward propagation step needs to be performed, so the rest of the process is Figure 1 Indicated by dotted line.

[0035] Step S1: Initialize the neural network, (refer to the traditional DSP algorithm) initialize the structure of the neural network to trainable variables (the structure of the network and the values ​​of the variables are initialized to the structure and values ​​of the traditional DSP);

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

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

[0038] Step S4: perform gradient calculation, calculate the gradient of the variable in the neural network with respect to the loss value obtained by calculating the loss value through the loss function;

[0039] Step S5: Perform back propagation, and use the calculated gradient to optimize the variable values ​​in the neural network through an optimization algorithm;

[0040] Step S6: Analyze the neural network variable values, extract the physical laws from the optimized neural network variables, and estimate the physical parameters.

[0041] Step S7: Reinitialize the neural network through the estimated physical parameters (refer to the traditional DSP algorithm again).

[0042] Specifically, for step S1:

[0043] Step S1.1: (The schematic diagram of initializing the neural network is as follows Figure 2As shown in FIG. 1 , the static linear operator in the DSP algorithm is initialized as a variable linear weight in the neural network, which is recorded as weight W (for example, the dispersion operator in the DBP method can be initialized to the linear weight of the neural network), thereby achieving the application of physical knowledge to the neural network;

[0044] Step S1.2: Initialize the static nonlinear operator in the DSP algorithm to a 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 of the neural network), so as to achieve the application of physical knowledge to the neural network;

[0045] Step S1.3: In the neural network initialized in this way, the combination of one or more linear weight modules followed by one or more nonlinear activation function modules is a layer of neural network (for example, in the classic DBP algorithm, the data is processed by alternating through a linear dispersion operator and a nonlinear operator. After it is initialized as a neural network, the combination of a dispersion operator and a nonlinear operator is a layer of neural network. The alternating passing through the two operators in the original DBP is changed to passing through multiple layers of neural networks one by one); the function of each layer of neural network is written as x=f(Wx0), where x is the output of the neural network of this layer, and x0 is the input of the neural network of this layer.

[0046] The value of DSP requires the physical parameter values ​​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 DSP, and then this set of values ​​can be used to initialize the variables of the neural network for the first time. Figure 2 In the figure, the static DSP represents the traditional channel equalization method, and its internal module is represented by a double-layer solid line; the variable neural network represents the proposed optical communication intelligent equalization algorithm based on physical regulation, and its internal module is represented by a three-layer solid line; both of them work in the way of taking the damaged signal as input and the equalized signal as output.

[0047] More specifically, for step S2, forward propagation is used to calculate the loss function during the training process (its output result may not be consistent with the actual original signal, and requires the cooperation of other processes in the optical communication intelligent equalization algorithm based on physical regulation); the output of forward propagation in the reasoning process represents its predicted actual original signal, thereby realizing compensation for signal damage (other processes in the optical communication intelligent equalization algorithm based on physical regulation are no longer required).

[0048] Furthermore, for step S3, the loss value calculated by the choice of loss function reflects the forward propagation result; the loss function is denoted as f L , which is chosen as the mean square error loss function; L = ||yy t || 2 / B,y tis the real original signal vector (corresponding to Figure 3 The original signal in the communication system becomes a damaged signal after passing through the communication system, which is recorded as y0. After forward propagation, it becomes y, which is the equalized signal vector obtained by forward propagation (corresponding to Figure 3 ), |||| represents the modulus operation on the vector, N is the length of the two signals, and L is the calculated loss value (y is calculated by y0 through the forward propagation of the multi-layer neural network, so L includes the role of multi-layer linear weights and activation functions, and L can be used to analyze the gradients of W and f in different layers).

[0049] Furthermore, for step S4, the gradient calculation step uses the loss value to calculate the gradient of the trainable variable value after the neural network is initialized with respect to the loss value. For a single-layer neural network (for example), its gradient needs to be calculated based on the derivative of the loss value L with respect to the weight W, expressed as:

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

[0051] Preferably, for step S6, the neural network variable value analysis process reads out the physical parameters in the original DSP algorithm corresponding to the neural network variables (such as the dispersion and nonlinear coefficient of the optical fiber), realizes the estimation of the physical parameters, and feedbacks the information from the neural network to the physics; when the neural network is initialized, if the variable of a single neural network corresponds to a single physical parameter, the variable optimized by back propagation is directly read out as the physical parameter; 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 parameter (because the optimizer roughly moves the loss value in a decreasing direction, the physical parameter corresponding to the optimized variable in the original DSP will be closer to the true value than before optimization).

[0052] Preferably, for step S7, the reinitialization process causes the optical communication intelligent balancing algorithm based on physical regulation to cycle multiple times, and an upper limit on the number of cycles of the training process is set to control the number of cycles, and the physical parameters used for reinitialization are closer to the true value than the last initialization (therefore, this will provide a better initialization for the optical communication intelligent balancing algorithm based on physical regulation, 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 "analyzing neural network variable values". Through the combination of the two, the neural network can obtain better initialization after each cycle, thereby achieving better performance. The optimization algorithm of traditional neural network often focuses on optimizing the value of the loss function, ignoring the internal physical laws, resulting in overfitting. The proposed optical communication intelligent equalization algorithm based on physical regulation can reduce the adverse effects of overfitting due to the existence of physical regulation.

[0054] The training process of optical communication intelligent equalization algorithm based on physical regulation when applied to optical communication system is as follows Figure 3 As shown, in order to carry out training, a known original signal needs to be input into the optical communication system, and a damaged signal is obtained after transmission; this damaged signal is input into the optical communication intelligent equalization algorithm based on physical adjustment that has been initialized for the first time, and then forward propagation is performed to obtain the output; the output of the neural network is compared with the known original signal, and the loss value is calculated, and then 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 real physical parameter, and then the neural network is initialized again according to the physical parameter, and the cycle of forward propagation, loss function calculation, gradient calculation, back propagation, variable value analysis, and reinitialization of the neural network is continued until the set cycle upper limit is reached.

[0055] The reasoning process of applying the optical communication intelligent equalization algorithm based on physical regulation to the optical communication system is as follows: Figure 4 As shown, the unknown original signal generates a damaged signal through the optical communication system. At this time, the optical communication intelligent equalization algorithm based on physical regulation has been trained, so it is only necessary to perform a single forward propagation of the damaged signal to obtain the equalized signal.

[0056] When the optical communication intelligent equalization algorithm based on physical regulation is applied to the optical fiber communication system, the optical fiber communication system parameters are set to 10×100km transmission of standard single-mode fiber (SMF), using quadrature amplitude modulation (QAM) format, fiber input power of 0dBm, optical signal-to-noise ratio of 15dB, transmission rate of 32Gbaud, and symmetric distributed Fourier method with a step size of 1km for simulation. In order to simulate the scenario of equalization under unknown optical fiber parameters, the initial parameters of DBP, LDBP and physical regulation neural network algorithms are all set to add perturbations to the parameters of SMF. The larger the perturbation, the worse the initialization of the algorithm and the more serious the mismatch with the actual situation. For the equalization algorithms based on neural networks (LDBP and physical regulation neural network), the forward and reverse propagation times are uniformly set to 650 times to ensure fair performance comparison.

[0057] The constellation diagram obtained by the optical communication intelligent equalization algorithm based on physical regulation is as follows Figure 5 As shown, taking the equalization of 64QAM signal as an example, the constellation diagram obtained by LDBP is Figure 5 Left, the corresponding bit error rate is 9.11×10 -2 , the constellation diagram obtained by the optical communication intelligent equalization algorithm based on physical regulation is Figure 5 Right, the corresponding bit error rate is 4.39×10 -3 ,It can be found that the balancing effect of the physical adjustment neural network is significantly better than the traditional LDBP method.

[0058] The bit error rate curve obtained by the optical communication intelligent equalization algorithm based on physical regulation is as follows: Figure 6 As shown in the figure, taking the equalization of 16QAM signal as an example, the bit error rate obtained by DBP, LDBP and physical adjustment neural network method is statistically analyzed as the β2 (dispersion) parameter mismatch changes at the time of initialization. It can be found that the static DBP method has the worst equalization ability for unknown systems, and the LDBP method has a certain adaptability due to the characteristics of its neural network. The proposed physical adjustment neural network equalization method combines neural network and physical adjustment, with the lowest bit error rate and the best effect.

[0059] The optical communication intelligent equalization algorithm based on physical regulation is used to estimate the physical parameters of the system. Figure 7 As the number of estimations increases, the physical parameters extracted by the algorithm (the blue solid line corresponding to Estimation) gradually converge from the parameters at the time of initialization to the true parameters (the red dotted line corresponding to Ground-truth), which reflects the ability of the algorithm to estimate physical parameters.

[0060] It is worth mentioning that the technical features such as the communication system involved in the patent application of this invention should be regarded as the prior art. The specific structure, working principle and possible control method and spatial layout method of these technical features can be selected by conventional methods in the field, and should not be regarded as the inventive point of the patent of this invention. The patent of this invention will not be further elaborated.

[0061] For those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An optical communication intelligent equalization algorithm based on physical regulation, characterized in that: The following steps are involved: Step S1: Initialize the neural network and initialize the structure of the neural network as trainable variables; Step S2: forward propagation is performed to equalize the damaged signal according to the current neural network variable value, wherein the input is the damaged signal transmitted through the optical communication system, and the output is the equalized signal, and the equalized signal output by the target is consistent with the original signal before the optical communication transmission; Step S3: Calculate the loss function, compare the equalized signal obtained by forward propagation with the real original signal, and calculate the loss value thereof through the loss function; Step S4: perform gradient calculation, calculate the gradient of the variable in the neural network with respect to the loss value obtained by calculating the loss value through the loss function; Step S5: Perform back propagation, and use the calculated gradient to optimize the variable values ​​in the neural network through an optimization algorithm; Step S6: Analyze the neural network variable values, extract the physical laws from the optimized neural network variables, and estimate the physical parameters. Step S7: Reinitialize the neural network using the estimated physical parameters.

2. According to claim 1, the optical communication intelligent equalization algorithm based on physical regulation is characterized in that: For step S1: Step S1.1: Initialize the static linear operator in the DSP algorithm to 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: Initialize the static nonlinear operator in the DSP algorithm to a variable nonlinear activation function in the neural network, denoted as function f, so as to achieve the application of physical knowledge to the neural network; Step S1.3: In the neural network initialized in this way, the combination of one or more linear weight modules followed by one or more nonlinear activation function modules forms a layer of neural network; the function of each layer of neural network is written as x=f(Wx0), where x is the output of this layer of neural network, and x0 is the input of this layer of neural network.

3. According to claim 2, the optical communication intelligent equalization algorithm based on physical regulation is characterized in that: For step S2, forward propagation is used to calculate the loss function during the training process; the output of forward propagation in the inference process represents the true original signal predicted by it, thereby compensating for signal damage.

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

5. According to claim 4, the optical communication intelligent equalization algorithm based on physical regulation is characterized in that: For step S4, the gradient calculation step uses the loss value to calculate the gradient of the trainable variable value after the neural network is initialized with respect to the loss value. For a single-layer neural network, its gradient needs to be calculated based on the derivative of the loss value L with respect to the weight W, expressed as:

6. The optical communication intelligent equalization algorithm based on physical regulation according to claim 5, characterized in that: For step S5, the variable optimization of the back-propagation process is implemented using the Adam optimizer.

7. The optical communication intelligent equalization algorithm based on physical regulation according to claim 6 is characterized in that: For step S6, the neural network variable value analysis process reads out the physical parameters in the original DSP algorithm corresponding to the neural network variables, realizes the estimation of the physical parameters, and feedbacks the information from the neural network to the physics; when the neural network is initialized, if the variable of a single neural network corresponds to a single physical parameter, the variable optimized by back propagation is directly read out as the physical parameter; 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 parameter.

8. The optical communication intelligent equalization algorithm based on physical regulation according to claim 7, characterized in that: For step S7, the reinitialization process causes the optical communication intelligent equalization algorithm based on physical regulation to cycle multiple times, and sets an upper limit on the cycle of the training process to control the number of cycles. The physical parameters used for reinitialization are closer to the true values ​​than those used in the previous initialization.

Images