Pump configuration method of optical fiber Raman amplifier with tunable gain and flatness

Abstract

The invention discloses a pumping configuration method of an optical fiber Raman amplifier with tunable gain and flatness, which belongs to the technical field of optical fiber Raman amplifiers, and comprises the following steps: establishing second-order Raman amplifier mathematical models with three structures; respectively establishing three different neural network models for the mathematical models of the second-order Raman amplifiers with the three structures, and performing reverse design of the system; and generating a test model by using the trained neural network model, and performing minimum pumping number configuration prediction under demand gain and flatness. According to the configuration method, the gain and the flatness of the optical fiber Raman amplifier can be guaranteed, meanwhile, the number of pump light is optimized to achieve maximization of cost effectiveness, adaptability is enhanced, and it is guaranteed that no matter how a user adjusts the target gain and the flatness, the optimal configuration of the minimum pump number can be provided; the cost of the Raman amplifier is effectively reduced, the operation efficiency and the response speed of the system are remarkably improved, optimal pump light parameter configuration under any gain requirement can be completed within several seconds, and the user experience and the actual application value of the system are greatly enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of optical fiber Raman amplifiers, and in particular relates to a pump configuration method for an optical fiber Raman amplifier with tunable gain and flatness. Background Art

[0002] With the commercialization of 5G technology and the rapid growth of communication business demand, optical fiber transmission systems are facing an urgent need for larger capacity and longer-distance relay-free transmission. In this context, traditional erbium-doped fiber amplifiers (EDFA) have been unable to meet the pursuit of broadband and high performance in modern optical communication systems due to their gain bandwidth limitations. Therefore, Raman amplifiers, with their unique advantages such as broadband amplification capability, low noise figure, and high saturation power, play an increasingly important role in all-optical communication systems.

[0003] Raman fiber amplifier (RFA) uses stimulated Raman scattering (SRS) effect to amplify signal light through the interaction between pump light and signal light. Compared with EDFA, RFA does not rely on rare earth element doping, and ordinary passive optical fiber can be used to generate gain, which makes RFA compatible with most communication systems. Communication transmission optical fiber can be multiplexed and can be used as both transmission medium and gain medium. In addition, RFA has great application potential in high-capacity mode division multiplexing optical communication systems due to its wavelength flexibility, gain spectrum width, fast response capability, and the use of reverse pumping to reduce noise coefficient.

[0004] In practical applications, the design of Raman amplifiers faces the challenges of fine-tuning gain and flatness, as well as how to reduce costs while ensuring performance. When dealing with these complex nonlinear problems, traditional design methods need to solve the Raman coupled wave differential equations that describe the interaction between pump light and signal light. This process is not only complex but also time-consuming. In addition, traditional numerical solution methods such as the Runge-Kutta method, the target method, and the average power method face non-convergence problems in some cases, which increases the complexity of the design. In order to address these challenges, researchers began to explore the use of machine learning algorithms, especially neural networks, to reverse design Raman amplifiers. This method predicts the pump configuration required to achieve a specific gain profile by learning the nonlinear mapping relationship between pump light parameters and gain characteristics to achieve the target gain curve. This method can not only quickly predict the pump configuration, but also effectively reduce the computational complexity and cost of the design process.

[0005] Inverse intelligent design technology has shown its outstanding advantages and great potential in the design and application of Raman amplifiers. However, past studies often ignored the cost considerations of the number of pumps when inversely designing the system. In practical applications, we should not only pursue the optimization of technical performance, but also take into account economic benefits, especially how to reduce costs by using the least number of Raman amplifiers while ensuring gain and flatness. This requires that when designing Raman amplifiers, the optimal balance between technical performance and economic benefits should be achieved to ensure that while meeting the performance requirements of the communication system, cost-effectiveness can be maximized. This design concept that comprehensively considers performance and cost is an important advancement in the field of Raman amplifier design. Summary of the invention

[0006] In view of the above-mentioned defects existing in the prior art, the present invention provides a pump configuration method for a fiber Raman amplifier with tunable gain and flatness. The configuration method can ensure the gain and flatness of the fiber Raman amplifier, and optimize the number of pump lights to maximize cost-effectiveness. It not only enhances adaptability, but also ensures that no matter how the user adjusts the target gain and flatness, the optimal configuration with the minimum number of pumps can be provided; it not only effectively reduces the cost of the Raman amplifier, but also significantly improves the operating efficiency and response speed of the system, and can complete the optimal pump light parameter configuration under any gain requirement within a few seconds, greatly enhancing the user experience and the practical application value of the system.

[0007] The present invention is achieved through the following technical solutions:

[0008] A pump configuration method for a fiber Raman amplifier with tunable gain and flatness, specifically comprising the following steps:

[0009] Step A: establishing mathematical models of second-order Raman amplifiers of three structures, namely a first structure model, a second structure model and a third structure model; wherein the first structure model includes a Raman amplifier with one second-order pump source and three first-order pump sources, the second structure model includes a Raman amplifier with one second-order pump source and four first-order pump sources, and the third structure model includes a Raman amplifier with one second-order pump source and five first-order pump sources;

[0010] Step B: Establish three different neural network models for the mathematical models of the second-order Raman amplifiers of the three structures respectively, and perform systematic reverse design, including:

[0011] Step B1: according to the set pump light parameter value range, the particle swarm optimization algorithm is used to optimize the pump light parameters, and the signal gain spectra corresponding to different pump parameters under different gain conditions are obtained during the optimization process;

[0012] Step B2: using the signal gain spectra under different gain conditions as input and the pump parameters under different gain conditions as output to form a new data set, and performing data processing;

[0013] Step B3: Use the new data set to train the neural network model to obtain a trained neural network model;

[0014] Step C: Generate a test model using the trained neural network model to predict the minimum number of pumps under the required gain and flatness.

[0015] Furthermore, in step B1, the mathematical models of the second-order Raman amplifiers of the three structures established in step A are solved by using the Runge-Kutta algorithm plus the shooting method, and then the flatness of the signal gain spectrum is optimized by iteratively adjusting the pump parameters in combination with the particle swarm optimization algorithm, as follows:

[0016] Firstly, the signal gain spectrum is obtained through the simplified nonlinear Raman coupling differential equation of the backward multi-pumped fiber Raman amplifier;

[0017] Then, at different gain levels, from m1db to m2db, with each interval of 2db as the target gain, the particle swarm optimization algorithm is used to find the minimum gain flatness under the target gain;

[0018] Finally, for each target gain, the same number of particles and iterations are set, so that the iterative process will generate small data sets with the same number under different gain spectra corresponding to different pump parameters.

[0019] Furthermore, the signal gain spectrum is obtained by using the simplified nonlinear Raman coupled differential equation of the optical fiber Raman amplifier, as follows:

[0020]

[0021] Where ± corresponds to the forward or reverse injection of the pump light source, P i , P j , P k represents the optical power of the i-th, j-th, and k-th channels respectively, and v i 、v j 、v k represents the optical frequency of the i-th, j-th, and k-th channels respectively, and g R (v i -v j ) represents the Raman gain coefficient between the two channels of light of the i-th and j-th channels, g R (v j -v k ) represents the Raman gain coefficient between the jth and kth channels, Γ represents the polarization factor, A eff Represents the effective core area of ​​the optical fiber, αj Represents the attenuation coefficient of the light wave of the jth channel transmitted in the optical fiber.

[0022] Further, in step B2, the details are as follows:

[0023] First, the small data sets (m2 / 2-m1 / 2) at different gain levels in step B1 are integrated into a large data set; this can prevent a large amount of data from overfitting the training of the MLP network and a small amount of data from underfitting the training of the MLP network;

[0024] Then, the signal gain spectrum of the large data set is used as input and the pump parameters of the large data set are used as output to generate the data set, and each set of data is arranged in ascending order of the pump wavelength, wherein the corresponding relationship of the power changes with the change of the pump wavelength;

[0025] Finally, outlier processing is performed on the arranged data set to obtain a new data set to improve the quality of the data set.

[0026] Further, in step B3, the details are as follows:

[0027] First, the new data set described in step B2 is normalized and divided into a training set and a test set in a ratio of 8:2;

[0028] Then, the hyperparameter optimization mechanism was added to train the neural network, with a maximum of 100 iterations; the search range of the hyperparameters was as follows: the number of layers was set to 2 to 6 layers, the number of neurons in each layer was set to 100 to 1000, and the learning rate was set to 10 -5 ~10 -1 , the dropout layer’s dropout rate range is set to 0 to 1; when the neural network iterates to the maximum number of times, the optimal hyperparameters suitable for the new data set are output;

[0029] Finally, the above optimal hyperparameters are input into a fully connected multilayer perceptron, and the neural network is trained using a 5-fold cross-validation method. Five trained neural network models are saved, and the average of all model prediction values ​​is taken to achieve the prediction output.

[0030] Furthermore, in step B3, the basic structure of the neural network model used is a fully connected multi-layer perceptron (MLP), a dropout layer is added between every two hidden layers to prevent overfitting, relu is used as the activation function, Adam is used as the optimizer, and BCE is used as the loss function.

[0031] Compared with the prior art, the advantages of the present invention are as follows:

[0032] The pump configuration method of a gain- and flatness-tunable optical fiber Raman amplifier of the present invention not only improves the flexibility of the system, but also ensures that when the user changes the target gain and flatness, the system can always provide the optimal pump configuration with the least number of pumps; it not only saves the cost of the Raman amplifier, but also improves the efficiency and response speed of the system, and can complete the optimal pump light parameter configuration under any gain within seconds, greatly improving the user experience and the application value of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0034] Figure 1 : A flow chart of a method for configuring a pump for a fiber Raman amplifier with tunable gain and flatness provided by an embodiment of the present invention;

[0035] Figure 2 : Structural diagram of optical fiber Raman amplifiers with three structures provided in the embodiments of the present invention;

[0036] Figure 3 : A flow chart of a particle swarm optimization algorithm provided by an embodiment of the present invention;

[0037] Figure 4 : A topological diagram of a neural network model provided by an embodiment of the present invention;

[0038] Figure 5 : A predicted output gain spectrum with a user demand gain of 12 dB and a flatness of 4 dB provided in an embodiment of the present invention;

[0039] Figure 6 : The predicted output gain spectrum for the user demand gain of 20dB and the flatness of 5dB provided in the embodiment of the present invention.

[0040] Figure 7 : The predicted output gain spectrum under the user demand gain of 19dB and flatness of 3dB provided in the embodiment of the present invention.

[0041] Figure 8 : The predicted output gain spectrum when the user demand gain is 27dB and the flatness is 0dB provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to clearly and completely describe the technical solution and its specific working process of the present invention, the specific implementation methods of the present invention are as follows in conjunction with the accompanying drawings of the specification:

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a pump configuration method for a fiber Raman amplifier with tunable gain and flatness, which specifically includes the following steps:

[0045] Step A: establishing mathematical models of second-order Raman amplifiers of three structures, namely a first structure model, a second structure model and a third structure model; wherein the first structure model includes a Raman amplifier with one second-order pump source and three first-order pump sources, the second structure model includes a Raman amplifier with one second-order pump source and four first-order pump sources, and the third structure model includes a Raman amplifier with one second-order pump source and five first-order pump sources;

[0046] Step B: Establish three different neural network models for the mathematical models of the second-order Raman amplifiers of the three structures respectively, and perform systematic reverse design, including:

[0047] Step B1: according to the set pump light parameter value range, the particle swarm optimization algorithm is used to optimize the pump light parameters, and the signal gain spectra corresponding to different pump parameters under different gain conditions are obtained during the optimization process;

[0048] Step B2: using the signal gain spectra under different gain conditions as input and the pump parameters under different gain conditions as output to form a new data set, and performing data processing;

[0049] Step B3: Use the new data set to train the neural network model to obtain a trained neural network model;

[0050] Step C: Generate a test model using the trained neural network model to predict the minimum number of pumps under the required gain and flatness.

[0051] The three pump prediction outputs are saved and input into the Raman coupling equation to calculate the corresponding signal gain spectrum. Check whether the gain and flatness of the signal gain spectrum meet the user's needs. If they do, the predicted number of pumps, corresponding pump parameters and predicted signal gain spectrum are output; if not, the three structures of Raman amplifiers each give a secondary option for the user to choose.

[0052] In this embodiment, step A is specifically as follows:

[0053] The mathematical models of the second-order Raman amplifiers of three structures used in the present invention are as follows: Figure 2As shown in (a), (b) and (c), m pump lights and n signal lights are coupled into the optical fiber through a wavelength division multiplexer for transmission, and the transmission direction of the pump light is opposite to that of the signal light. The backward pumping method can effectively reduce the time and distance of interaction between the signal light and the pump light, and reduce the system noise. In order to achieve long-distance transmission of distributed fiber Raman amplifiers, 80km of G.652 optical fiber is used as the gain medium to complete the amplification of the C+L band signal light.

[0054] This embodiment establishes three mathematical models of second-order Raman amplifiers, namely a first structural model, a second structural model and a third structural model; wherein the first structural model includes a Raman amplifier with 1 second-order pump source and 3 first-order pump sources, the second structural model includes a Raman amplifier with 1 second-order pump source and 4 first-order pump sources, and the third structural model includes a Raman amplifier with 1 second-order pump source and 5 first-order pump sources; 87 signal lights are amplified, i.e., m=4, 5, 6, and n=87. Figure 2 In the example, Input means input, Output means output, and optical fiber is used as the transmission medium.

[0055] For the distributed fiber Raman amplifier, its amplification process is to use the stimulated Raman scattering effect generated between lights of different wavelengths in the optical fiber to achieve the amplification of the signal light during transmission. By controlling the wavelengths of the signal light and the pump light during transmission, the two are within a specific frequency difference range, allowing the signal light to fully absorb the energy of the pump light to achieve its own amplification. The mathematical model of the Raman fiber amplifier is a simplified nonlinear Raman coupling equation, as shown in equation (1). In the present invention, the main considerations are the loss of the signal light and the pump light during their own transmission, and the stimulated Raman scattering effect between lights of different wavelengths (signal light and signal light, signal light and pump light, pump light and pump light);

[0056]

[0057] Where ± corresponds to the forward or reverse injection of the pump light source, P i , P j , P k represents the optical power of the i-th, j-th, and k-th channels respectively, and v i 、v j 、v k represent the optical frequencies of the i, j, and k channels respectively, and g R (v i -v j ) represents the Raman gain coefficient between the two channels of light of the i-th and j-th channels, g R (v j -v k ) represents the Raman gain coefficient between the jth and kth channels, Γ represents the polarization factor, A effRepresents the effective core area of ​​the optical fiber, α j Represents the attenuation coefficient of the light wave of the jth channel transmitted in the optical fiber.

[0058] There are two main parameters for evaluating the performance of Raman amplifiers, namely the output gain and gain flatness of the fiber Raman amplifier. The switching gain and gain flatness are defined as equations (2) and (3), respectively:

[0059]

[0060] ripple=max(G onoff )-min(G onoff ) (3)

[0061] In formula (2), P outon , P outoff are the output optical power of the signal light when the pump is turned on and the output optical power of the signal light when the pump is turned off. Formula (3) is the difference between the maximum switch gain value and the minimum switch gain value, which is expressed by the gain flatness ripple. The smaller the difference is, the more consistent the output gain value of the signal light after amplification is, and the flatter the output gain curve is. The above two are important indicators for evaluating the performance of fiber Raman amplifiers.

[0062] According to the nonlinear Raman coupling equation, the approximate pump light parameter value range (wavelength and power) that affects the gain performance of the fiber Raman amplifier is determined. Then, according to the set pump light parameter value range, the particle swarm optimization algorithm is used to optimize the pump light parameters, and the data set generated during the optimization process under different gain conditions is obtained. The specific parameter settings are as follows: the signal light power is 0.1mW, the signal light interval is 1.2nm, and the signal light of 1523.6nm-1626.8nm in the C+L band is amplified, a total of 87 channels. The parameter range settings of the three structures of the pump light sources used are as follows:

[0063] For the first structural model: λ1: 1350nm~1370nm, p1: 0.4W~1.5W, λ2: 1420nm~1450nm, p2: 0.05W~0.6W, λ3: 1440nm~1480nm, p3: 0.05W~0.6W, λ4: 1480nm~1510nm, p4: 0.01W~0.4W;

[0064] For the second structural model, λ1: 1350nm~1370nm, p1: 0.4W~1.4W, λ2: 1420nm~1450nm, p2: 0.05W~0.6W, λ3: 1435nm~1470nm, p3: 0.05W~0.6W, λ4: 1460nm~1480nm, p4: 0.05W~0.6W,; λ5: 1500nm~1510nm, p5: 0.01W~0.2W;

[0065] For the third structural model: λ1: 1350nm~1370nm, p1: 0.4W~1.4W, λ2: 1420nm~1440nm, p2: 0.05W~0.6W, λ3: 1435nm~1460nm, p3: 0.05W~0.6W, λ4: 1440nm~1470nm, p4: 0.05W~0.6W, λ5: 1460nm~1480nm, p5: 0.05W~0.6W, λ6: 1500nm~1510nm, p6: 0.01W~0.2W;

[0066] use Figure 3 The traditional particle swarm algorithm shown in FIG. 1 can improve the quality of the data set by optimizing the generated data set within a given pump wavelength and power range. The specific particle swarm optimization algorithm part optimizes the RFA mainly to reduce the gain flatness ripple while meeting the target output gain G, so that the signal output gain spectrum is as smooth as possible under a given switch gain. This embodiment uses the traditional particle swarm algorithm to optimize G and ripple. The specific algorithm process includes the following steps: First, set the search dimension to D, the total number of particles to M, and the total number of iterations to maxiteration. In the D-dimensional search space, the position information of each particle is X. i =(x i1 ,x i2 ,...,x iD ), the speed information is Vi=(v i1 ,v i2 ,...,v iD ). Where i represents the i-th particle in the search space, x i1 ,x i2 ,...,x iD are the position information of each component, v i1 ,v i2 ,...,v iD are the velocity information of each component respectively. The fitness function of the particle is

[0067] fit=max{||max(G onoff -G t )||,||Gt -min(G onoff )||, max(G onoff ) - min(G onoff )}, where G t is the target gain, which is set at intervals of 2 dB between 10 and 30 dB. For application in the design of an optical fiber Raman amplifier, the velocity and position information of the particles in this algorithm correspond to the wavelength and power information of the pump light. The particle optimization process is a process of finding extreme values, including two extreme values. One is the individual extreme value of each particle, which is particlesbestposition, and the other is the global extreme value gobalbestposition after comparison of each particle in the entire search space. According to equations (4) and (5)

[0068]

[0069] the velocity information and position information of the particles are updated at different iteration times. In the formula, k is the iteration number; d = 1, 2,..., D; w is a non - negative constant, called the inertia weight; c 1，2 is a non - negative constant, called the learning factor; r 1,2 is a random number in the interval [0, 1]. w can be understood as the influence of the motion inertia on the particles, c1 represents the influence of the historical experience of a single particle on itself, and c2 represents the decision - making influence generated by the historical experience of other particles. Continuously update the iteration number and determine whether the maximum iteration number is reached. If i = maxiteration, the iteration ends and the optimal position and velocity information are output. If i < maxiteration, then continue to repeat the previous step. Set the number of particles to 20. After reaching the maximum iteration number of 100, the algorithm stops. The iteration process includes the signal gain spectra corresponding to different pump light parameters. Take the signal gain spectra as the input quantity and the pump light parameters as the output quantity to form a data set (20×100 groups). Finally, the Raman amplifier data sets of the three structures all contain the same number of data, and the data set size is 10×20×100 = 20000 groups. Then sort this data set in ascending order of the pump wavelength, and keep the original corresponding relationship of the pump power. Then screen the outliers of the data set by the quartile method to obtain a new data set with improved quality.

[0070] In this embodiment, steps B1 and B2 are specifically as follows:

[0071] Normally, directly integrating the differential equations of the Raman coupled waves is a very complex task, so numerical methods are usually used to solve them, such as the Runge-Kutta method and the average power method. For backward-pumped Raman amplifiers, it is usually necessary to combine the targeting method with the Runge-Kutta method or the average power method to solve their numerical solutions. However, in some cases, these methods may face convergence problems. The present invention adopts a neural network to reversely design the Raman amplifier, and constructs a neural network model through training, replacing the traditional numerical solution process, which greatly saves the time for solving nonlinear equations and improves computational efficiency. In theory, when the neural network training is successful, the optimal pump light parameter configuration under any gain spectrum can be obtained in only seconds.

[0072] The present invention adopts the MLP network model, and its topology diagram is as follows: Figure 4 As shown in Figure 1, the neural network model consists of four parts: input layer, hidden layer, dropout layer and output layer. n ] is the input of the neural network, which represents the switching gain spectrum of n signal lights of different wavelengths after amplification, K = [λ1,…,λ m ,P1,…,P m ] is the output of the neural network, and is the wavelength λ and power P of the m pump lights. Different neural network structure algorithms will lead to different training results. The optimal hyperparameter determination process of the MLP model is as follows:

[0073] According to the data set, the input layer is assigned 87 variables and the output layer corresponds to three types of Raman amplifiers, namely 8, 10, and 12 variables. The signal gain spectrum is used as the input of the neural network model, and the pump light parameters are used as the output of the neural network model. The new data set after sorting and outlier screening in step A is normalized and randomly divided into training set and test set in a ratio of 8:2. Taking the first structural model as an example, MLP sets relu as the activation function, and a dropout layer is added between every two hidden layers to prevent overfitting. The sigmoid function is added to the output layer, and the loss function is BCE. TPE and MedianPruner hyperparameter optimization mechanisms are added to optimize the hyperparameters of the MLP model. The maximum number of iterations is 100. The search space of the hyperparameter optimization algorithm is: the number of hidden layers ranges from 2 to 6, the number of neurons in each layer ranges from 100 to 1000, and the learning rate ranges from 10 -5 ~10 -1 The dropout rate of the dropout layer ranges from 0.1 to 0.5. The Adam algorithm is used to train it, and the training loss is used as the comparison quantity to compare the performance of the model under different hyperparameters.

[0074] Finally, the MLP hyperparameters of the first structural model are determined as follows: 5 hidden layers, the number of neurons in the hidden layers are: 1004, 614, 981, 750, 893, the drop rate is: 0.12086524246966658, and the learning rate is: 6.784520422066518×10 -5 .

[0075] The MLP hyperparameters of the second structural model are: 4 hidden layers, the number of neurons in the hidden layers are: 768, 717, 805, 736, the drop rate is: 0.1486468166880957, and the learning rate is: 1.8085185680019987×10 -4 .

[0076] The MLP hyperparameters of the third structure model are: 4 hidden layers, the number of neurons in the hidden layers are: 1010, 542, 553, 718, the drop rate is 0.1056348376220844, and the learning rate is 8.890363718003893×10 -5 .

[0077] In this embodiment, step B3 is specifically as follows:

[0078] The optimal hyperparameters obtained in step B2 are input into the MLP model, and the model is retrained using the 5-fold cross-validation method. Then, the average of the five models is used to verify the model performance on the test set. K-fold cross-validation can improve the accuracy and reliability of model evaluation. By dividing the data set into K subsets of equal size and repeatedly using these subsets during training and testing, more stable and reliable model evaluation results can be obtained. This method can avoid the deviation caused by the division of a single training set and test set, and provide more accurate performance evaluation. At the same time, K-fold cross-validation can also reduce overfitting: by training and evaluating the model on multiple training sets and validation sets, it helps to detect and control overfitting. By using different combinations of training sets and validation sets, K-fold cross-validation can better capture the generalization ability of the model, thereby reducing the risk of overfitting. After retraining, MSE and R2 values ​​are used as evaluation indicators of the model, where MSE represents the mean square error between the predicted value and the actual value. When the MSE value is closer to 0, the error is smaller. The R value represents the closeness between the predicted value and the actual value. When the R value is closer to 1, the prediction accuracy of the neural network model is higher. After training, the MSE of the MLP of the first structural model is 0.8448, and the R2 is 0.9616; the MSE of the MLP of the second structural model is 0.6607, and the R2 is 0.9544; the MSE of the MLP of the third structural model is 0.6841, and the R2 is 0.9578.

[0079] After determining the MLP model, verify that the error of the MLP model in predicting the pump light parameters is reasonable, and then the MLP model can be directly called. The trained MLP models of the three structures of the Raman amplifier are integrated together to form a test model. When the user inputs the gain value and flatness of the required signal gain spectrum, the test model can randomly generate a set of signal gain spectrum values ​​that strictly meet the input conditions, and then call the three MLP models to get the predicted pump light parameter configurations of the three structures.

[0080] In this embodiment, step C is specifically as follows:

[0081] The pump light parameter configurations of the three structures are used as input, and the actual gain curve is obtained by solving the Raman coupling equation. By comparing the gain values ​​and flatness between the target gain curve and the three actual gain curves, the pump configuration with the minimum number of pumps that meets the target gain and flatness is screened out and then output. If there is no pump configuration that meets the target gain and flatness, the system outputs the three closest secondary pump configuration solutions given by the test model. The following are 4 examples of some response changes of the test model when user needs change.

[0082] Figure 5 The output gain spectrum corresponding to the final pump configuration given by the test model is shown when the user enters the target gain value: 12dB, and the target flatness value is 4dB. The final predicted pump configuration given by the model is: minimum 4 pump sources, 4 pump wavelengths are: 1361.0nm, 1430.8nm, 1465.1nm, 1507.8nm; 4 pump powers are: 0.7584W, 0.1625W, 0.1458W, 0.0964W. The gain mean and gain flatness of the signal gain spectrum corresponding to this pump configuration are 12.1075dB and 2.9795dB respectively.

[0083] Figure 6 The output gain spectrum corresponding to the final pump configuration given by the test model is shown when the user enters the target gain value: 20dB, and the target flatness value is 5dB. The final predicted pump configuration given by the model is: a minimum of 5 pump sources, 5 pump wavelengths are: 1353.5nm, 1423.7nm, 1445.1nm, 1470.6nm, 1507.9nm; 5 pump powers are: 1.0693W, 0.2973W, 0.1607W, 0.1448W, 0.0469W. The gain mean and gain flatness of the signal gain spectrum corresponding to this pump configuration are 13.09dB and 3.2dB respectively.

[0084] Figure 7The output gain spectrum corresponding to the final pump configuration given by the test model is shown when the user enters the target gain value: 19dB, and the target flatness value is 3dB. The final predicted pump configuration given by the model is: minimum 6 pumps, 6 pump wavelengths are: 1357.0nm, 1426.5nm, 1436.0nm, 1451.6nm, 1473.8nm, 1507.3nm; 6 pump powers are: 0.8662W, 0.2194W, 0.1964W, 0.1219W, 0.1047W, 0.0601W. The gain mean and gain flatness of the signal gain spectrum corresponding to this pump configuration are 19.6265dB and 2.8287dB respectively.

[0085] Figure 8 The following figure shows the output gain spectrum corresponding to the final pump configuration given by the test model when the user enters the target gain value: 27dB and the target flatness value 0dB. Since no pump configuration can meet the flatness of 0dB, the test model gives three alternatives. Alternative 1: The mean and flatness of the gain spectrum of the four pumps are: mean: 26.9924dB, flatness: 8.4556dB. The four pump wavelengths are: 1355.3nm, 1428.4nm, 1459.3nm, 1505.0nm; the four pump powers are: 1.2757W, 0.5074W, 0.4164W, 0.0636W. Alternative 2: The mean and flatness of the gain spectrum of the five pumps are: mean: 26.4766dB, flatness: 5.2722dB. The five pump wavelengths are: 1357.1nm, 1425.8nm, 1443.5nm, 1468.2nm, 1507.3nm; the five pump powers are: 1.2412W, 0.4690W, 0.2542W, 0.2022W, 0.0374W. Secondary option 3: The mean and flatness of the gain spectrum of the six pumps are: mean: 27.4119dB, flatness: 3.8447dB. The six pump wavelengths are: 1354.8nm, 1423.9nm, 1439.5nm, 1452.8nm, 1471.2nm, 1507.9nm; the six pump powers are: 1.1720W, 0.4785W, 0.2287W, 0.2362W, 0.1327W, 0.0296W.

[0086] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.

[0087] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0088] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A pump configuration method for a fiber Raman amplifier with tunable gain and flatness, characterized in that: The specific steps include: Step A: establishing mathematical models of second-order Raman amplifiers of three structures, namely a first structure model, a second structure model and a third structure model; wherein the first structure model includes a Raman amplifier with one second-order pump source and three first-order pump sources, the second structure model includes a Raman amplifier with one second-order pump source and four first-order pump sources, and the third structure model includes a Raman amplifier with one second-order pump source and five first-order pump sources; Step B: Establish three different neural network models for the mathematical models of the second-order Raman amplifiers of the three structures respectively, and perform systematic reverse design, including: Step B1: according to the set pump light parameter value range, the particle swarm optimization algorithm is used to optimize the pump light parameters, and the signal gain spectra corresponding to different pump parameters under different gain conditions are obtained during the optimization process; Step B2: using the signal gain spectra under different gain conditions as input and the pump parameters under different gain conditions as output to form a new data set, and performing data processing; Step B3: Use the new data set to train the neural network model to obtain a trained neural network model; Step C: Generate a test model using the trained neural network model to predict the minimum number of pumps under the required gain and flatness.

2. The pump configuration method of a fiber Raman amplifier with tunable gain and flatness as claimed in claim 1, characterized in that: In step B1, the mathematical models of the second-order Raman amplifiers of the three structures established in step A are solved by using the Runge-Kutta algorithm plus the shooting method, and then the flatness of the signal gain spectrum is optimized by iterative pumping parameters in combination with the particle swarm optimization algorithm, as follows: Firstly, the signal gain spectrum is obtained through the simplified nonlinear Raman coupling differential equation of the backward multi-pumped fiber Raman amplifier; Then, at different gain levels, from m1db to m2db, with each interval of 2db as the target gain, the particle swarm optimization algorithm is used to find the minimum gain flatness under the target gain; Finally, for each target gain, the same number of particles and iterations are set, so that the iterative process will generate small data sets with the same number under different gain spectra corresponding to different pump parameters.

3. The pump configuration method of a fiber Raman amplifier with tunable gain and flatness as claimed in claim 2, characterized in that: The signal gain spectrum is obtained by using the simplified nonlinear Raman coupled differential equation of the optical fiber Raman amplifier, as follows: Where ± corresponds to the forward or reverse injection of the pump light source, P i , P j , P k represents the optical power of the i-th, j-th, and k-th channels respectively, and v i 、v j 、v k represents the optical frequency of the i-th, j-th, and k-th channels respectively, and g R (v i -v j ) represents the Raman gain coefficient between the two channels of light of the i-th and j-th channels, g R (v j -v k ) represents the Raman gain coefficient between the jth and kth channels, Γ represents the polarization factor, A eff Represents the effective core area of ​​the optical fiber, α j Represents the attenuation coefficient of the light wave of the jth channel transmitted in the optical fiber.

4. The pump configuration method of a fiber Raman amplifier with tunable gain and flatness as claimed in claim 1, characterized in that: In step B2, the details are as follows: First, the small data sets (m2 / 2-m1 / 2) at different gain levels in step B1 are integrated into a large data set; this can prevent a large amount of data from overfitting the training of the MLP network and a small amount of data from underfitting the training of the MLP network; Then, the signal gain spectrum of the large data set is used as input, and the pump parameters of the large data set are used as output to generate the data set, and each set of data is arranged in ascending order of the pump wavelength, wherein the corresponding relationship of the power changes with the change of the pump wavelength; Finally, outlier processing is performed on the arranged data set to obtain a new data set to improve the quality of the data set.

5. The pump configuration method of a fiber Raman amplifier with tunable gain and flatness as claimed in claim 1, characterized in that: In step B3, the details are as follows: First, the new data set described in step B2 is normalized and divided into a training set and a test set in a ratio of 8:2; Then, TPE and MedianPruner hyperparameter optimization mechanisms were added to train the neural network, with a maximum of 100 iterations. The search range of hyperparameters was as follows: the number of layers was set to 2 to 6 layers, the number of neurons in each layer was set to 100 to 1000, and the learning rate was set to 10. -5 ~10 -1 , the dropout layer’s dropout rate range is set to 0.1 to 0.5; when the neural network iterates to the maximum number of times, the optimal hyperparameters suitable for the new data set are output; Finally, the above optimal hyperparameters are input into a fully connected multilayer perceptron, and the neural network is trained using a 5-fold cross-validation method. Five trained neural network models are saved, and the average of all model prediction values ​​is taken to achieve the prediction output.

6. The pump configuration method of a fiber Raman amplifier with tunable gain and flatness as claimed in claim 5, characterized in that: In step B3, the basic structure of the neural network model used is a fully connected multi-layer perceptron. A dropout layer is added between every two hidden layers to prevent overfitting. Relu is the activation function, Adam is the optimizer, and BCE is the loss function.

Images