CN115270641A

CN115270641A - Reverse design method and system for backward multi-pump Raman fiber amplifier

Info

Publication number: CN115270641A
Application number: CN202211022246.XA
Authority: CN
Inventors: 巩稼民; 卢姣姣; 高睿杰; 郭刘飞; 吴成超; 刘威; 裴梦杰
Original assignee: Xian University of Posts and Telecommunications
Current assignee: Xian University of Posts and Telecommunications
Priority date: 2022-08-25
Filing date: 2022-08-25
Publication date: 2022-11-01

Abstract

The invention relates to a backward multi-pump Raman fiber amplifier reverse design method and a system, which belong to the field of Raman fiber amplifiers. The improved particle swarm optimization algorithm is combined with the neural network, the accuracy of the neural network model is improved, meanwhile, the Raman fiber amplifier for amplifying the C + L waveband signal light is obtained, the calculation efficiency is improved, and meanwhile, the output gain value and the output gain flatness of the Raman amplifier are improved.

Description

Reverse design method and system for backward multi-pump Raman fiber amplifier

Technical Field

The invention relates to the field of Raman fiber amplifiers, in particular to a reverse design method and a reverse design system of a backward multi-pump Raman fiber amplifier.

Background

In the field of optical communication research, an optical Fiber Amplifier directly affects the transmission performance of an optical Fiber transmission system as an indispensable core device, and a Raman Fiber Amplifier (RFA) plays an important role in a 6G transmission communication system in the future because of its outstanding advantages of large bandwidth, good gain flatness, low noise figure, good compatibility, etc., and is gradually a hot spot of research in the field of optical amplifiers.

Theoretically, when a proper pump light parameter configuration is selected in the RFA design process, the stimulated Raman scattering effect can be utilized to realize the Raman amplification of the full waveband, and when the amplified signal power is close to the pump power, the optical gain is only reduced by 3dB. The main challenge of RFA design is to select the pump power and wavelength to generate a specific gain profile, which requires solving the differential equation of raman coupled wave describing the nonlinear effect between the pump light and the signal light transmitted along the fiber, and only approximate solutions can be found due to the complexity of the equation, and the current methods for numerically solving the equation are the langousta method, the average power method, and the targeting method, and the method for describing the multi-channel optical wave interaction propagation equation together using genetic algorithm, which are complicated and in some cases face the situation of non-convergence.

In recent years, most of research on raman amplifiers focuses on how to configure optimization of pump power and wavelength, and in order to solve the problem, genetic algorithms, artificial bee colony algorithms, differential evolution algorithms and the like are provided, although flat gain required by RFA can be achieved, in the running process, the parameters of the algorithms need to be adjusted according to different problems to perform running iteration again, and in the design process, all parts of the algorithms related to solving the raman coupling wave equation are time-consuming, and repeated circulation causes unnecessary work repetition and time loss.

In order to improve the solving process of the nonlinear differential equation comprehensively, in recent years, a solving process of learning the nonlinear mapping relation between the pumping light and the signal light by adopting a neural network algorithm in machine learning and training a neural network model to replace a Raman coupled wave equation is proposed. Chen statics et al, university of Fuzhou in 2018, combined Extreme Learning Machine (ELM) and differential evolution algorithm (DE), and through the rapid learning speed and high generalization of ELM and the strong global search capability of DE, the final gain ripple is less than 0.5dB. Darko Zibar et al, denmark, 2020 proposed a reverse design of Raman amplifiers using machine learning fine optimization to make accurate numerical predictions of the pump design for any Raman gain curve. The machine learning framework of raman amplifiers that can be used to design and model any gain was experimentally characterized by niara c.de Moura et al, denmark, 2021, using different fiber types for testing, and finally showed errors below 0.5dB. All the above can show that it is feasible and effective to reverse design the raman amplifier by using the neural network, but none of them improves the performance and amplification bandwidth of the raman fiber amplifier while considering improving the efficiency and reducing the error.

Disclosure of Invention

The invention aims to provide a reverse design method and a reverse design system for a backward multi-pump Raman fiber amplifier, which are used for improving the output gain value and the output gain flatness of the Raman amplifier while greatly improving the calculation efficiency.

In order to achieve the purpose, the invention provides the following scheme:

a backward multi-pump raman fiber amplifier reverse design method, the method comprising:

determining pump optical parameters influencing the gain performance of the backward multi-pump Raman fiber amplifier;

optimizing the pump optical parameters by adopting an improved particle swarm optimization algorithm according to the set value range of the pump optical parameters to obtain the optimal pump optical parameters under different gain conditions;

calculating a gain curve under each gain condition by utilizing a nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier according to the optimal pump optical parameter under each gain condition;

forming a data set by taking gain curves under different gain conditions as input quantities and taking optimal pump light parameters under different gain conditions as output quantities;

training a BP neural network model by using the data set to obtain a trained BP neural network model;

and inputting the target gain curve into the trained BP neural network model, and outputting the optimal pump light parameters of the backward multi-pump Raman fiber amplifier.

Optionally, the determining a pump optical parameter affecting the gain performance of the backward multi-pump raman fiber amplifier specifically includes:

the simplified nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier is established as

Wherein + -corresponds to forward or reverse injection of the pump light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j, k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the i-th and j-th channel lights _R (v _j -v _k ) Denotes the Raman gain coefficient, K, between the j-th and K-th channel two lights _eff Denotes the polarization factor, A _eff Representing the effective core area, alpha, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Represents RayleighThe scattering coefficient, K and h represent the Boltzmann constant and the Planckian constant, respectively,

is a bose-einstein factor, and T is the absolute temperature of the optical fiber;

determining pump optical parameters influencing the gain performance of the backward multi-pump Raman amplifier according to the nonlinear Raman coupling differential equation; the pump light parameters include power and wavelength of the pump light source.

Optionally, the weights in the improved particle swarm optimization algorithm are iteratively updated according to the following formula:

w＝(w ₁ +w ₂ )·(T'-In)/T'+w ₂

wherein w represents a weighting factor at the current iteration number, w ₁ And w ₂ Respectively representing the initial value and the final value of the weight factor, wherein In represents the current iteration number, in is less than or equal to T ', and T' represents the total iteration number.

Optionally, training a BP neural network model by using the data set to obtain a trained BP neural network model specifically includes:

setting a BP neural network model to comprise an input layer, a hidden layer and an output layer;

on the premise that the number of preset hidden layers is 3-8 and each layer comprises 1-100 neurons, 600 BP neural network models are constructed;

and carrying out deletion completion, abnormal value processing and normalization processing on the data set, and carrying out 7:1.5:1.5, randomly dividing the processed data set into a training set, a testing set and a verification set;

training 600 BP neural network models by adopting a TRAINSCG algorithm according to the training set and the verification set;

comparing the performances of 600 trained BP neural network models by using a test set and taking a regression value R value and a mean square error MSE value as evaluation indexes, and determining that the BP neural network model with the optimal performance has a structure of 6 hidden layers and each layer comprises 33 neurons;

and training the BP neural network model with the optimal performance by adopting the data set to obtain the trained BP neural network model.

Optionally, the value range of the pump light parameter is as follows: the wavelength range of each pump light is 1400-1500nm, and the power value range of each pump light source is 0-2W.

A backward multi-pumped raman fiber amplifier reverse design system, the system comprising:

the influence factor determining module is used for determining pump optical parameters influencing the gain performance of the backward multi-pump Raman fiber amplifier;

the optimizing module is used for optimizing the pump optical parameters by adopting an improved particle swarm optimization algorithm according to the set value range of the pump optical parameters to obtain the optimal pump optical parameters under different gain conditions;

the gain calculation module is used for calculating a gain curve under each gain condition by utilizing a nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier according to the optimal pump optical parameter under each gain condition;

the data set forming module is used for taking the gain curves under different gain conditions as input and taking the optimal pump light parameters under different gain conditions as tags to form a data set;

the training module is used for training a BP neural network model by using the data set to obtain a trained BP neural network model;

and the application module is used for inputting the target gain curve into the trained BP neural network model and outputting the optimal pump light parameters of the backward multi-pump Raman fiber amplifier.

Optionally, the influence factor determining module specifically includes:

a differential equation establishing submodule for establishing a simplified nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier as

Wherein + -corresponds to forward or reverse injection of the pump light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j and k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the i-th and j-th channel lights _R (v _j -v _k ) Denotes the Raman gain coefficient, K, between the j-th and K-th channel two lights _eff Denotes the polarization factor, A _eff Represents the effective core area, α, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Expressing Rayleigh scattering coefficients, K and h expressing Boltzmann constants and Planck constants, respectively,

the pumping optical parameter determining submodule is used for determining pumping optical parameters influencing the gain performance of the backward multi-pumping Raman amplifier according to the nonlinear Raman coupling differential equation; the pump light parameters include power and wavelength of the pump light source.

Optionally, the weight in the improved particle swarm optimization algorithm is iteratively updated according to the following formula:

w＝(w ₁ +w ₂ )·(T'-In)/T'+w ₂

where w represents a weighting factor at the current iteration number, w ₁ And w ₂ And respectively representing the initial value and the final value of the weight factor, wherein In represents the current iteration number, in is less than or equal to T ', and T' represents the total iteration number.

Optionally, the training module specifically includes:

the model setting submodule is used for setting a BP neural network model to comprise an input layer, a hidden layer and an output layer;

the model construction submodule is used for constructing 600 BP neural network models on the premise that the number of the preset hidden layer layers is 3-8 and the number of each layer containing 1-100 neurons;

a partitioning submodule, configured to perform deficiency completion, outlier processing, and normalization processing on the data set, and perform, according to 7:1.5:1.5, randomly dividing the processed data set into a training set, a testing set and a verification set;

the training submodule is used for training 600 BP neural network models by adopting a TRAINSCG algorithm according to the training set and the verification set;

the test submodule is used for comparing the performances of 600 trained BP neural network models by using a test set and taking a regression value R value and a mean square error MSE value as evaluation indexes, and determining that the BP neural network model with the optimal performance has a structure of 6 hidden layers and each layer comprises 33 neurons;

and the model training submodule is used for training the BP neural network model with the optimal performance by adopting the data set to obtain the trained BP neural network model.

Optionally, the value range of the pump light parameter is: the wavelength range of each pump light is 1400-1500nm, and the power value range of each pump light source is 0-2W.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

the invention discloses a backward multi-pump Raman fiber amplifier reverse design method and a backward multi-pump Raman fiber amplifier reverse design system. The improved particle swarm optimization algorithm is combined with the neural network, the accuracy of the neural network model is improved, meanwhile, the Raman fiber amplifier for amplifying the C + L waveband signal light is obtained, the calculation efficiency is improved, and meanwhile, the output gain value and the output gain flatness of the Raman amplifier are improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

Fig. 1 is a flowchart of a reverse design method of a backward multi-pump raman fiber amplifier according to an embodiment of the present invention;

fig. 2 is a structural diagram of a backward multi-pump raman amplifier according to an embodiment of the present invention;

FIG. 3 is a flow chart of an improved particle swarm optimization algorithm provided by an embodiment of the present invention;

FIG. 4 is a schematic diagram of a neural network construction provided by an embodiment of the present invention;

FIG. 5 is a diagram of a neural network model topology provided by an embodiment of the present invention;

FIG. 6 is a graph of RFA output gain before and after optimization according to an embodiment of the present invention;

FIG. 7 is a distribution diagram of R values for different hidden layers and different numbers of neurons according to an embodiment of the present invention; fig. 7 (a) is a distribution diagram of the R value at the number of hidden layers of 3, fig. 7 (b) is a distribution diagram of the R value at the number of hidden layers of 4, fig. 7 (c) is a distribution diagram of the R value at the number of hidden layers of 5, fig. 7 (d) is a distribution diagram of the R value at the number of hidden layers of 6, fig. 7 (e) is a distribution diagram of the R value at the number of hidden layers of 7, and fig. 7 (f) is a distribution diagram of the R value at the number of hidden layers of 8;

FIG. 8 is a graph of a distribution of MSE values at different hidden layers and different numbers of neurons according to an embodiment of the present invention; fig. 8 (a) is a distribution diagram of the MSE values at 3 hidden layer numbers, fig. 8 (b) is a distribution diagram of the MSE values at 4 hidden layer numbers, fig. 8 (c) is a distribution diagram of the MSE values at 5 hidden layer numbers, fig. 8 (d) is a distribution diagram of the MSE values at 6 hidden layer numbers, fig. 8 (e) is a distribution diagram of the MSE values at 7 hidden layer numbers, and fig. 8 (f) is a distribution diagram of the MSE values at 8 hidden layer numbers;

fig. 9 is a graph of the variation of the R value and the MSE value of the optimal model with the number of hidden layers according to the embodiment of the present invention; fig. 9 (a) is a graph of the variation of the R value with the number of hidden layers, and fig. 9 (b) is a graph of the variation of the MSE value with the number of hidden layers;

FIG. 10 is a graph of the R value and the error result under the optimal neural network model structure provided by the embodiment of the present invention; fig. 10 (a) is a graph of an R value and an error training result in an optimal neural network model structure, (b) in fig. 10 is a graph of an R value and an error verification result in the optimal neural network model structure, and (c) in fig. 10 is a graph of an R value and an error test result in the optimal neural network model structure;

fig. 11 is a graph of target gain and predicted gain at different gain values according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention more comprehensible, the present invention is described in detail with reference to the accompanying drawings and the detailed description thereof.

The embodiment of the invention provides a reverse design method of a backward multi-pump Raman fiber amplifier, which comprises the following steps as shown in figure 1:

s1, determining pump optical parameters influencing the gain performance of the backward multi-pump Raman fiber amplifier.

The structure of the backward multi-Pump RFA structure adopted by the invention is shown in figure 2, wherein m beams of Pump light (Pump light) and n beams of Signal light (Signal light) are respectively coupled into an Optical fiber through an Optical Multiplexer Unit (OMU) for transmission,and the transmission direction of the pump light is opposite to that of the signal light. The backward pumping mode can effectively reduce the interaction time and distance between the signal light and the pumping light and reduce the system noise. In order to realize high-gain output of RFA, tellurite fiber is adopted as a gain medium, and amplified signal light is output after redundant pump light is filtered by a Filter (Filter), so that amplification of C + L waveband signal light is completed. In the present invention, for example, four pump lights are used to amplify 125 signal lights, i.e., m =4, n =125. In fig. 2, input represents Input, transmission represents Transmission, output represents Output, and the wavelengths of the n signal beams are λ ₁ 、λ ₂ 、λ ₃ 、…、λ _n The wavelengths of the m beams of pump light are respectively lambda _p1 、λ _p2 、λ _p3 、…、λ _pm 。

For RFA, the amplification process of the RFA is to realize the amplification of signal light in the transmission process by utilizing the stimulated Raman scattering effect generated between light beams with different wavelengths in optical fibers, and the signal light and the pump light are in a specific frequency difference range by controlling the wavelengths of the signal light and the pump light in the transmission process, so that the signal light fully absorbs the energy of the pump light to realize the amplification of the RFA. The mathematical model of the Raman fiber amplifier is a simplified nonlinear Raman coupling differential equation shown in formula (1). In the invention, the loss of signal light during self transmission, the stimulated Raman scattering effect between light beams with different wavelengths (signal light and signal light, signal light and pump light, and pump light), and the influence of spontaneous radiation, rayleigh scattering and the like on RFA are mainly considered.

In the formula, the +/-corresponds to the forward or reverse injection of the pumping light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j and k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the i-th and j-th channel lights _R (v _j -v _k ) Denotes the Raman gain coefficient between the j-th and K-th channel lights, K _eff Denotes the polarization factor, A _eff Represents the effective core area, α, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Expressing Rayleigh scattering coefficients, K and h expressing Boltzmann constants and Planck constants, respectively,

is the Bose-Einstein factor, and T is the absolute temperature of the fiber.

The main parameters for evaluating the performance of a raman amplifier are two aspects, namely the output gain and the gain fluctuation of the RFA. And the gain flatness thereof are defined as formulas (2) and (3), respectively:

Δ＝max(G)-min(G) (3)

p in formula (2) _j (0)、P _j And (L) is the initial input optical power of the jth path of signal light and the optical power after the transmission distance of L respectively. Equation (3) is the difference between the maximum gain value and the minimum gain value, and is expressed by the gain flatness Δ, and the smaller the difference, that is, the more uniform the output gain values of the amplified signal light becomes, the flatter the output gain curve becomes. Both of these are important indicators for evaluating RFA performance.

Determining pump optical parameters influencing the gain performance of the backward multi-pump Raman amplifier according to a nonlinear Raman coupling differential equation; the pump light parameters include the power and wavelength of the pump light source.

And S2, optimizing the pump optical parameters by adopting an improved particle swarm optimization algorithm according to the set value range of the pump optical parameters to obtain the optimal pump optical parameters under different gain conditions.

The method limits the weight on the basis of the traditional particle swarm algorithm, improves the global search capability, avoids the search particles from tending to regionalization, and obtains the optimal solution by optimizing in the given pumping wavelength and power range.

The particle swarm optimization algorithm part is specifically improved:

the optimization of RFA mainly comprises the steps of improving the output gain G of the system and reducing the gain flatness delta to enable the gain output to be a smooth curve. The improved particle swarm algorithm is adopted to optimize G and delta.

The Improved particle swarm optimization (IPSO optimization) is also derived from simulation of bird foraging behavior, and is different from the traditional particle swarm optimization in that the weight factor is Improved, so that the particles are optimized by changing inertia in a search range. Referring to fig. 3, the specific flow of the algorithm includes the following steps:

firstly, setting a search dimension as D, the total number of particles as M and the total number of iteration times as T.

1) In the D-dimensional search space, the position information of each particle is X _i ＝(x _i1 ,x _i2 ,…,x _iD ) Velocity information is V _i ＝(v _i1 ,v _i2 ,…,v _iD ). Where i denotes the ith particle in the search space, x _i1 ,x _i2 ,…,x _iD Position information, v, for each component, respectively _i1 ,v _i2 ,…,v _iD Respectively, velocity information for each component. The fitness function of the particle is fit (·) =1/Δ. For the design applied to RFA, the velocity and position information of the particles in the algorithm corresponds to the wavelength and power information of the pump light. The particle optimization process is a process for finding an extremum, and comprises two extremums, one is that the individual extremum of each particle is p _id The other is the global extremum p of each particle in the whole search space after comparison _gd 。

2) And (5) updating the weight factor, the speed information and the position information of the particles at different iteration times according to the formula (4).

Wherein w ₁ And w ₂ Respectively is an initial value and a final value of the weight factor w, and In is less than or equal to TThe number of previous iterations. The latter two formulas represent the velocity information v and the position information x of the (k + 1) th particle in the (d) th iteration respectively. c. C ₁ 、c ₂ Is a learning factor, r ₁ 、r ₂ Is [0, 1]]Random numbers within a range.

Continuously updating the iteration times, judging whether the maximum iteration times is reached, if In = T ', ending the iteration, outputting the optimal position and speed information, and if In < T', continuously repeating the previous step.

And S3, calculating a gain curve under each gain condition by utilizing a nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier according to the optimal pump optical parameters under each gain condition.

And S4, taking the gain curves under different gain conditions as input quantities, and taking the optimal pump light parameters under different gain conditions as output quantities to form a data set.

The IPSO algorithm is adopted to optimize the power and the wavelength of the pump light, 3000 groups of optimal pump light parameter configuration sets are formed to form a data set after 3000 iterations, and the size of the data set is 9 x 3000.

And S5, training the BP neural network model by using the data set to obtain the trained BP neural network model.

Generally speaking, directly integrating the raman coupled wave differential equation is a very complicated process, and generally a numerical solution method such as the longgutta method and the average power method is adopted, while for the backward pumping RFA, an analytic solution of the target method and the longgutta method or the average power method is required to be obtained by combining the target method and the longgutta method, and in some cases, the result is subject to non-convergence. The method adopts a mode of reverse design of the neural network RFA, constructs a neural network model through multiple times of training, replaces the traditional numerical analysis process, greatly saves the time for solving the nonlinear equation, and improves the calculation efficiency.

Equation (1) is simplified to a function of Y = f (X), where X represents the input data of equation (1), i.e. the power and wavelength information of the pump light, Y represents the output value of equation (1), i.e. the output gain of the RFA, and the function f (·) is regarded as a complex non-linear mapping of the whole X to Y. Reverse designThe RFA is targeted to obtain a matching X value based on a given Y value, i.e. f needs to be determined ^-1 The mapping relation of the function (DEG), in order to solve the problem, a neural network model NN (DEG) is constructed to learn the complex mapping. Theoretically, when NN (-) training is successful, the optimal pump optical parameter configuration under any gain can be obtained only by the second order.

And taking the gain value as an input variable of the neural network model, and taking the pump light parameter as an output variable of the neural network model. And (3) performing deletion completion, abnormal value processing and normalization processing on 3000 groups of samples, and then performing the following steps of 7:1.5: the scale of 1.5 was randomly divided into a training set, a test set, and a validation set.

The neural network model comprises three parts, namely an input layer, a hidden layer and an output layer, wherein the input layer is allocated to be one variable and the output layer is allocated to be eight variables according to data set data. The number of layers of the hidden layers is set to be 3-8 in an experiment, 600 models are constructed by the number of 1-100 neurons in each layer, the models are trained by the aid of the TRAINSCG algorithm, regression values R and mean square error MSE serve as evaluation indexes, performances of different models are compared, and finally the structure of the neural network is determined to be 6 hidden layers and each layer contains 33 neurons. And after the neural network model is determined, the data set is used for training the neural network model, the reasonable error of the neural network model for predicting the pump light parameters is verified, and then the neural network model can be directly called.

The establishment of the neural network model NN (-) is divided into three steps, establishment, training and verification. As shown in fig. 4, firstly, an Improved Particle Swarm Optimization (IPSO) is used to optimize a parameter K of pump light, where the IPSO is referred to as a reference, and is used as an input of an equation solver, a raman coupled wave differential equation is solved to obtain a gain spectrum G, and G and K together construct a data set, a nonlinear mapping from G to K is learned by constructing a neural network NN (), and once a neural network model NN (-) is trained successfully, a target gain curve is input, a predicted pump light parameter configuration can be obtained, the pump light parameter configuration is used as an input, an actual gain curve is obtained by solving the equation, and accuracy of prediction by using the model is determined by comparing errors between the target gain curve and the actual gain curve. In fig. 4, actual Gain spectrum represents Actual Gain spectrum, equation solver represents Equation solver, raman coupled wave Equation represents solution of coupled wave Equation, gain represents Gain, frequency represents Frequency, data set represents Data set, IPSO algorithm represents IPSO algorithm, target Gain spectrum represents Target Gain spectrum, neural Networks represents Neural network, and Predicted Gain spectrum represents Predicted Gain peak.

When the absolute value difference between the two is closer to 0, the network model is high in accuracy, the Raman coupled wave equation can be solved in an accurate replacement of the traditional mode, and if the absolute value difference between the two is too large, the Raman coupled wave equation is continuously optimized until the error is within a feasible range.

The invention adopts a BP neural network model, the topological diagram of which is shown in figure 5, and the neural network model is divided into an Input Layer (Input Layer), a Hidden Layer (Hidden Layer) and an Output Layer (Output Layer). G = [ G ] ₁ ,G ₂ ,G ₃ ,…,G _n ] ^T For the input of the neural network, the gain values of the amplified output of the signal lights with n different wavelengths are represented, and K = [ lambda ] ₁ ,…,λ _m ，p ₁ ,…,p _m ] ^T The output of the neural network is the wavelength λ and power p of the m pump lights. Wherein C is _x Number of hidden layers, R, included in the structure _y The number of the neurons contained in each hidden layer unit is represented, different neural network structure algorithms lead to different training results, and the optimal neural network model structure is determined in the following process:

the power of signal light is set to be 0.01W, the interval of the signal light is set to be 0.8nm, and the signal light with the C + L wave band of 1530nm-1630nm is amplified for 125 paths. The amplification of the signal light is realized by adopting four pumping light sources together, and the optimization effectiveness of the IPSO algorithm is explained firstly in the invention.

And setting parameters and an optimization range of the IPSO algorithm, and outputting optimal pump light parameter configuration after iteration is finished in a specified search range. As shown in fig. 6, for comparison graphs of RFA output gain spectra before and after optimization, where empirical values of 1400nm, 1430nm, 1440nm, and 1470nm are adopted for pump light parameters before optimization, and powers are 0.01W, 0.04W, 0.14W, and 0.8W, respectively, the obtained output gain is 15.82dB, the flatness is 1.51dB, and compared with the optimized output gain value of 15.93dB, and the gain fluctuation range of 0.36dB, it can be seen that the RFA flatness after using the IPSO algorithm is greatly improved, that is, the optimization algorithm has higher adaptability to the RFA pump light parameters optimization, and a data set is constructed based on the improved flatness.

The method comprises the steps of optimizing power and wavelength of pump light by adopting an IPSO algorithm, forming 3000 groups of optimal pump light parameter configuration sets to form a data set after 3000 iterations, randomly dividing the data set into a training set, a verification set and a test set according to the proportion of 70%,15% and 15%, training a network model by adopting a time-saving TRAINSCG function, and taking a Mean Square Error (MSE) and a regression R value as evaluation indexes of network performance, wherein the MSE represents a mean square error between a predicted value and an actual value, when the MSE value is closer to 0, the smaller the error is, the R value represents the tightness between the predicted value and the actual value, and when the R value is closer to 1, the higher the prediction accuracy of a neural network model is. In the process of training the neural network model, the neural network structure, such as the number of hidden layers and the number of neurons in each hidden layer, needs to be determined. In order to simplify simulation, the number of the neural networks contained in each hidden layer is set to be the same and is in the range of (0, 100), the condition that the number of the hidden layers is too large and too many hidden layers are abandoned, the number of the hidden layers is set to be in the range of [3,8], and 600 groups of neural network models with different structures are constructed.

As shown in fig. 7 and 8, the distribution of R and MSE values with hidden layer and neuron numbers under 600 models, respectively. When the number of hidden layers is increased to the 5 th layer and the 6 th layer, the aliasing is reduced, the overall R value is distributed between [0.94, 1], and the MSE value is distributed in the range of (0, 1 ]. The effect is better when the number of the neurons is in the range of [20,80 ].

Comparing the optimal R value and the MSE value of each layer, as shown in fig. 9, (a) and (b) are respectively the maximum R value and the minimum MSE value of each layer with the hidden layer number in the range of [3,8], the overall R value gradually increases and the MSE value gradually decreases as the hidden layer number increases, but when the hidden layer number increases to the 7 th layer and the 8 th layer, the overall performance decreases, which indicates that the learning ability has a deviation when the neural network model is gradually complex. And it can be clearly seen that the overall energy level of the neural network model is better when the hidden layer is at layer 6.

The optimal neural network structure finally determined for the inverse design RFA contains 6 hidden layers, each layer containing 33 neurons. On the basis of this configuration, 3000 sets of data were trained, and the results are shown in fig. 10. And (a), (b) and (c) are respectively training, verifying and testing results, regression R values are respectively 0.9964, 0.99513 and 0.99541, all data points are concentrated on a straight line and are distributed, the linear fitting capacity is high, and the target and the output are approximately in a linear relation.

To check the error between the target value and the predicted value, the gain is arbitrarily specified, and the wavelength and power value of the pump light matched therewith are predicted by NN (·) model and are used as input to compare with the actual gain value obtained by solving using a numerical solver, as shown in fig. 11. For the error distribution between the target and the actual under different gains, the dotted line in the figure is the target gain value, the solid line is the gain value obtained by prediction of the neural network model, the predicted values are respectively set to 4dB, 7.65dB, 10.75dB and 13.55dB, it can be seen that when the target gain value is smaller, the error between the target value and the predicted value is smaller, the superposition degree of the two lines is higher, because when the power value of the pump light is low, the degree of the stimulated Raman scattering effect between the pump light and the signal light is smaller, the complexity is low, the error is smaller, and the error is not more than 0.47dB at all.

And S6, inputting the target gain curve into the trained BP neural network model, and outputting the optimal pump light parameters of the backward multi-pump Raman fiber amplifier.

In order to overcome the defects of slow response, narrow bandwidth and the like of a next-generation communication super-large-capacity and super-high-speed optical transmission system, the application discloses an effective method for quickly predicting and optimizing the performance of a backward multi-pump Raman optical fiber amplifier. Firstly, the parameter configuration of pump light is optimized by adopting an improved particle swarm optimization algorithm, then a Raman fiber amplifier is reversely designed by combining a neural network algorithm to learn the nonlinear mapping relation between output gain and pump light parameters, and the target output gain is determined so as to generate pump wavelength and power matched with the target output gain, thereby replacing the traditional method for solving a Raman coupled wave differential equation by numerical values. The two methods are combined, so that the accuracy of the neural network model is improved, and the Raman fiber amplifier for amplifying the C + L waveband signal light is obtained. The experimental result shows that when the designed Raman fiber amplifier is used for prediction, the error between the target value and the predicted value is not more than 0.47dB. The method has certain reference significance for the flexible design of the future Raman fiber amplifier.

The invention adopts an improved particle swarm optimization algorithm to optimize the wavelength and the gain of the pump light, limits the weight on the basis of the traditional particle swarm optimization algorithm, improves the global search capability, avoids the tendency of searching particles to regionalization, obtains the optimal solution by optimizing in a given pump wavelength and power range, takes the optimal solution as a data set, reversely designs a backward multi-pump RFA through a neural network algorithm, learns the nonlinear mapping relation between an RFA target gain curve and the pump light wavelength and power by utilizing a neural network, effectively solves the complex calculation process of Raman coupled wave equation integral in designing the Raman amplifier, greatly improves the calculation efficiency, improves the output gain value of the Raman amplifier and slows down the fluctuation of the output gain. The discrete Raman fiber amplifier facing the C + L waveband is designed and obtained by combining two algorithms, and the Raman amplifier obtained by design can provide any gain by determining the optimal neural network model and the parameters of the optimization algorithm, has higher precision and improves the flatness of output gain. An effective mode is provided for designing RFA with specific gain, and meanwhile, the method provides theoretical basis and technical support for research of the optical transmission method of the 6G communication network.

The embodiment of the invention also provides a backward multi-pump Raman fiber amplifier reverse design system, which comprises:

the data set forming module is used for forming a data set by taking the gain curves under different gain conditions as input and taking the optimal pump light parameters under different gain conditions as tags;

The influence factor determining module specifically comprises:

a differential equation establishing submodule for establishing a simplified nonlinear Raman coupling differential equation of the backward multi-pump Raman fiber amplifier

In the formula, the +/-corresponds to the forward or reverse injection of the pumping light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j and k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the i-th and j-th channel lights _R (v _j -v _k ) Denotes the Raman gain coefficient between the j-th and K-th channel lights, K _eff Denotes the polarization factor, A _eff Representing the effective core area, alpha, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Expressing Rayleigh scattering coefficients, K and h expressing Boltzmann constants and Planck constants, respectively,

the pumping light parameter determining submodule is used for determining pumping light parameters influencing the gain performance of the backward multi-pumping Raman amplifier according to the nonlinear Raman coupling differential equation; the pump light parameters include power and wavelength of the pump light source.

The weight in the improved particle swarm optimization algorithm is iteratively updated according to the following formula:

w＝(w ₁ +w ₂ )·(T'-In)/T'+w ₂

where w represents a weighting factor at the current iteration number, w ₁ And w ₂ Respectively representing the initial value and the final value of the weight factor, wherein In represents the current iteration number, in is less than or equal to T ', and T' represents the total iteration number.

The training module specifically comprises:

the model setting submodule is used for setting the BP neural network model to comprise an input layer, a hidden layer and an output layer;

a partitioning submodule, configured to perform deficiency completion, outlier processing, and normalization processing on the data set, and perform the following processing according to 7:1.5:1.5, randomly dividing the processed data set into a training set, a testing set and a verification set;

the training submodule is used for training the 600 BP neural network models by adopting a TRAINSCG algorithm according to the training set and the verification set;

The value range of the pump light parameter is as follows: the wavelength range of each pump light is 1400-1500nm, and the power value range of each pump light source is 0-2W.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the description of the method part.

The principle and the implementation mode of the invention are explained by applying a specific example, and the description of the embodiment is only used for helping to understand the method and the core idea of the invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.

Claims

1. A backward multi-pump raman fiber amplifier reverse design method, comprising:

optimizing the pump light parameters by adopting an improved particle swarm optimization algorithm according to the set value range of the pump light parameters to obtain the optimal pump light parameters under different gain conditions;

2. The method according to claim 1, wherein the determining of the pump light parameters affecting the gain performance of the backward multi-pump raman fiber amplifier comprises:

Wherein + -corresponds to forward or reverse injection of the pump light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j and k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the i-th and j-th channel lights _R (v _j -v _k ) Denotes the Raman gain coefficient, K, between the j-th and K-th channel two lights _eff Denotes the polarization factor, A _eff Representing the effective core area, alpha, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Expressing Rayleigh scattering coefficients, K and h expressing Boltzmann constants and Planck constants, respectively,

3. The method of claim 1, wherein the weights in the improved particle swarm optimization algorithm are iteratively updated according to the following formula:

w＝(w ₁ +w ₂ )·(T'-In)/T'+w ₂

4. The method according to claim 1, wherein the training of the BP neural network model using the data set to obtain the trained BP neural network model specifically comprises:

5. The method of claim 1, wherein the pump light parameter ranges from: the wavelength range of each pump light is 1400-1500nm, and the power value range of each pump light source is 0-2W.

6. A backward multi-pump raman fiber amplifier reverse engineering system, comprising:

7. The system of claim 6, wherein the influencing factor determining module specifically comprises:

Wherein + -corresponds to forward or reverse injection of the pump light source, P _i 、P _j 、P _k Respectively representing the optical power of the i, j, k channels, v _i 、v _j 、v _k Respectively represent the optical frequencies of the i, j and k channels, g _R (v _i -v _j ) Denotes a Raman gain coefficient, g, between the two lights of the ith and jth channels _R (v _j -v _k ) Denotes the Raman gain coefficient, K, between the j-th and K-th channel two lights _eff Denotes the polarization factor, A _eff Represents the effective core area, α, of the optical fiber _j Representing the attenuation coefficient, gamma, of the optical wave of the j-th channel transmitted in the optical fiber _j Expressing the Rayleigh scattering coefficient, K and h expressing the Boltzmann constant and the Planck constant, respectively,

8. The system of claim 6, wherein the weights in the improved particle swarm optimization algorithm are iteratively updated according to the following formula:

w＝(w ₁ +w ₂ )·(T'-In)/T'+w ₂

wherein w represents a weighting factor at the current iteration number, w ₁ And w ₂ And respectively representing the initial value and the final value of the weight factor, wherein In represents the current iteration number, in is less than or equal to T ', and T' represents the total iteration number.

9. The system of claim 6, wherein the training module specifically comprises:

the model construction submodule is used for constructing 600 BP neural network models on the premise that the number of preset hidden layers is 3-8 and the number of neurons in each layer is 1-100;

10. The system of claim 6, wherein the pump light parameter ranges from: the wavelength range of each pump light is 1400-1500nm, and the power value range of each pump light source is 0-2W.