A fiber Raman amplifier gain adaptive control method based on two-stage neural network

Abstract

The present invention relates to a method for adaptive gain control of a fiber Raman amplifier based on a two-stage neural network, comprising the following steps: obtaining training data; determining a two-stage neural network topology structure based on the training data; training the two-stage neural network using the training data; calculating the pump light power and wavelength at the amplifier's target gain value using the trained first-stage neural network; calculating a predicted amplifier gain value at the pump light power and wavelength output by the first-stage neural network using the trained second-stage neural network; subtracting the predicted gain value obtained by the second-stage neural network from the target gain value to obtain a gain error; and cyclically using a gradient descent method to update the neuron link weights based on the gain error, thereby improving the accuracy of the fiber Raman amplifier's output gain value. The present invention uses the first-stage neural network to calculate the pump light power and wavelength at the amplifier's target gain value, uses the second-stage neural network to calculate the amplifier's predicted gain value, and uses the gradient descent method to update the neuron link weights based on the gain error between the amplifier's predicted gain value and the target gain value, ultimately obtaining a highly accurate amplifier output gain value and the corresponding pump light power and wavelength values.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber Raman amplifier gain adaptive control, and in particular to an optical fiber Raman amplifier gain adaptive control method based on a neural network. Background Art

[0002] In long-distance, ultra-high-speed submarine cable networks, optical amplifiers are required to compensate for the power loss of optical signals during link transmission. Therefore, optical amplifiers are key components for ensuring the normal operation of long-distance, ultra-high-speed submarine cable networks. Among various optical amplifiers, broadband fiber Raman amplifiers based on multi-wavelength pumping technology are considered to be helpful for the realization of long-distance, ultra-high-speed submarine cable networks. They have low noise and a wide bandwidth gain spectrum. Moreover, by appropriately adjusting the wavelength and power of the pump light, a composite gain spectrum with target fluctuations can be generated within any desired wavelength range.

[0003] Fiber Raman amplifiers (FRA) in submarine cable networks face two operational challenges. First, signals in different channels of multi-band systems are subject to non-uniform effects such as Kerr nonlinearity, amplified spontaneous emission noise, and nonlinear scattering. To achieve the highest information transmission rate and longest transmission distance, the FRA gain spectrum must exhibit a non-flat waveform. Furthermore, when the signal power in the system changes, the FRA gain spectrum must be adjusted accordingly. Second, with the rise of the underwater Internet of Things (IoT) and the concept of submarine cable communication technology supporting marine scientific observations, coupled with the development of various wireless, wired, acoustic, and electromagnetic access methods, such as short-range underwater acoustic communication, underwater wireless optical communication, and underwater pluggable connectors, submarine cable networks will evolve from closed, non-scalable networks to ones that support flexible and dynamic networking, thereby improving their accessibility and scalability. Signal light of any wavelength band may be added or removed from the system at any time, requiring the relay system to adjust pump light parameters based on the number and wavelength of signal light in the system to achieve optimal gain spectrum coverage. In summary, facing the reality that the parameters such as the number, wavelength and power of transmitted signal light in the submarine optical cable transmission link are subject to sudden changes; in order to expand the transmission distance of the submarine optical cable network and improve the system transmission capacity, it is necessary to provide the submarine optical cable network with a broadband optical fiber Raman amplifier with adaptive and adjustable gain spectrum.

[0004] With the widespread attention that artificial intelligence (AI) has garnered in academia and industry, neural network technology has rapidly developed, and its application in fiber-optic communication systems has also gained attention. A trained neural network can calculate the required fiber Raman amplifier pump parameters based on the target gain spectrum, demonstrating that neural network-based design methods are an effective way to achieve adaptive gain spectrum adjustment in fiber Raman amplifiers. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] The present invention proposes a design method for a fiber Raman amplifier based on a double-stage neural network, the purpose of which is to achieve a high-accuracy adaptive adjustment function of the gain spectrum of the fiber Raman amplifier.

[0007] (2) Technical solution

[0008] To solve the above technical problems, the present invention proposes a fiber Raman amplifier gain adaptive control method based on a two-stage neural network, comprising:

[0009] Acquiring training data, wherein the data set includes pump light wavelength data, pump light power data, and fiber Raman amplifier gain value;

[0010] Determining the topological structures of the primary neural network and the secondary neural network based on the training data to obtain initial primary neural network and initial secondary neural network models; the primary neural network uses the fiber Raman amplifier gain value as an input layer and the pump light wavelength and pump light power as an output layer; the secondary neural network uses the pump light wavelength and pump light power as an input layer and the fiber Raman amplifier gain value as an output layer;

[0011] Using the training data to train the primary neural network model and the secondary neural network model to obtain trained primary neural network model and secondary neural network model;

[0012] Calculating the pump light power value and pump light wavelength value required at the target gain value of the Raman amplifier according to the trained first-level neural network model;

[0013] Calculating the predicted gain value of the fiber Raman amplifier under the pump light power value and pump light wavelength value output by the first-level neural network according to the trained second-level neural network model;

[0014] Subtracting the predicted gain value obtained by the secondary neural network from the target gain value of the fiber Raman amplifier to obtain a gain error;

[0015] According to the gain error combined with the gradient descent method, the weight values of each neuron link of the first-level neural network and the second-level neural network are updated, the gain error between the predicted gain value and the target gain value is calculated again, and the gain error and the gradient descent method are cyclically used to adjust the neuron link weight value, and finally a high-accuracy amplifier output gain value and the corresponding pump light power and wavelength values are obtained.

[0016] As a preferred example, before determining the topological structures of the primary neural network and the secondary neural network models according to the training data, the method further includes:

[0017] The training data is subjected to missing data completion, outlier processing, and normalization processing.

[0018] As a preferred example, determining the topological structure of the primary neural network according to the training data to obtain an initial primary neural network model specifically includes:

[0019] Determining an input layer of a primary neural network model according to the fiber Raman amplifier gain value in the training data;

[0020] Determining a hidden layer of a primary neural network model according to a gain error of the optical fiber Raman amplifier;

[0021] The output layer of the primary neural network model is determined according to the wavelength data of the pump light and the power data of the pump light in the training data.

[0022] As a preferred example, determining the topological structure of the secondary neural network based on the training data to obtain an initial secondary neural network model specifically includes:

[0023] Determining an input layer of a secondary neural network model according to wavelength data of the pump light and power data of the pump light in the training data;

[0024] Determining a hidden layer of a secondary neural network model according to a gain error of the optical fiber Raman amplifier;

[0025] The output layer of the secondary neural network model is determined according to the fiber Raman amplifier gain value in the training data.

[0026] As a preferred example, the input layer of the primary neural network contains 40 neurons, the number of hidden layers is 3 and each layer contains 100 neurons, and the output layer contains 6 neurons; the input layer of the secondary neural network contains 6 neurons, the number of hidden layers is 3 and each layer contains 100 neurons, and the output layer contains 40 neurons;

[0027] As a preferred example, the input layer, hidden layer and output layer in the primary neural network and the secondary neural network are connected in sequence using an S-type transfer function sigmoid.

[0028] (3) Beneficial effects

[0029] The present invention discloses a method for designing a fiber Raman amplifier based on a two-stage neural network. The topology and connection weights of a primary neural network and a secondary neural network are determined based on training data. The trained primary neural network is used to calculate the pump light power and pump light wavelength at the target gain value of the Raman amplifier. The trained secondary neural network model is used to calculate the predicted gain value of the fiber Raman amplifier at the pump light power and pump light wavelength output by the primary neural network. The predicted gain value obtained by the secondary neural network is subtracted from the target gain value of the fiber Raman amplifier to obtain a gain error. The gradient descent method is then used cyclically to update the neuron connection weights based on the gain error, thereby improving the accuracy of the fiber Raman amplifier output gain value. The present invention uses the primary neural network to calculate the pump light power and wavelength at the target gain value of the amplifier, the secondary neural network to calculate the predicted gain value of the amplifier, and the gradient descent method is used to update the neuron connection weights based on the gain error between the predicted gain value and the target gain value, ultimately obtaining a highly accurate amplifier output gain value and the corresponding pump light power and wavelength values. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 This is a flow chart of the fiber Raman amplifier design method based on a two-stage neural network;

[0032] Figure 2 It is a structural diagram of a two-stage neural network model;

[0033] Figure 3 This is a comparison chart of the simulation results of the fiber Raman amplifier gain spectrum accuracy before and after optimization. DETAILED DESCRIPTION

[0034] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] The purpose of the present invention is to provide a method for calculating the gain value of an optical fiber Raman amplifier based on a two-stage neural network, so as to realize adaptive control of the gain of the optical fiber Raman amplifier.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the technical solutions of the present invention are further described in detail below with reference to the accompanying drawings and specific embodiments:

[0037] like Figure 1 As shown, a fiber Raman amplifier gain adaptive control method based on a two-stage neural network includes:

[0038] Step 101: Acquire training data, where the data set includes pump light wavelength data, pump light power data, and corresponding fiber Raman amplifier gain values.

[0039] According to theory and experiments, pump light power and pump light wavelength are the main factors affecting the gain value of the fiber Raman amplifier. Therefore, in order to achieve adaptive adjustment of the gain spectrum of the fiber Raman amplifier, it is necessary to realize adaptive control of the pump light power and pump light wavelength of the fiber Raman amplifier.

[0040] Step 102: Determine the topological structures of the primary neural network and the secondary neural network based on the training data to obtain initial primary neural network and secondary neural network models. Step 102 specifically includes:

[0041] Step 1021: Determine the input layer of the first-level neural network based on the fiber Raman amplifier gain value in the training data, and determine the output layer of the first-level neural network model based on the pump light wavelength data and pump light power data in the training data.

[0042] Step 1022: Determine the input layer of the secondary neural network model according to the pump light wavelength data and the pump light power data in the training data, and determine the output layer of the secondary neural network according to the fiber Raman amplifier gain value in the training data.

[0043] Step 1023: Determine the initial first-level neural network and the second-level neural network hidden layer based on the training data.

[0044] Among them, the two-stage neural network model structure is as follows Figure 2 As shown, it consists of a first-level neural network and a second-level neural network, where the input layer of the first-level neural network contains 40 neurons, G1(λ s1 )…G 40 (λ s40 ) represents the gain value of 40 different wavelength signal lights; the output layer of the first-level neural network contains 6 neurons, P p1 …P p3 Represents three pump light power values, λ p1 …λ p3represents three pump light wavelength values; the number of hidden layers of the first-level neural network is 3 and each layer contains 100 neurons. The input layer of the second-level neural network is the output layer of the first-level neural network; the output layer of the second-level neural network contains 40 neurons, G1(λ s1 )…G 40 (λ s40 ) represents the gain values of 40 different wavelengths of signal light. The second-level neural network has three hidden layers, each containing 100 neurons. The input, hidden, and output layers of the first and second-level neural networks are sequentially connected using a sigmoid transfer function.

[0045] Step 103: Use the training data to train the primary neural network model and the secondary neural network model respectively to obtain a trained primary neural network and a trained secondary neural network.

[0046] Step 104: Calculate the pump light power and pump light wavelength at the target gain value of the Raman amplifier using the trained first-level neural network model;

[0047] Step 105: Calculate the fiber Raman amplifier gain value at the pump light power and pump light wavelength output by the first-level neural network using the trained second-level neural network model;

[0048] Step 106: Subtracting the gain prediction value obtained by the secondary neural network from the target gain value of the fiber Raman amplifier to obtain a gain error;

[0049] Step 107: Update the weight values of each neuron link of the first-level neural network and the second-level neural network according to the gain error combined with the gradient descent method, recalculate the gain error between the predicted gain value and the target gain value, and cyclically use the gain error and the gradient descent method to adjust the neuron link weight value, and finally obtain a high-accuracy amplifier output gain value and the corresponding pump light power and wavelength values.

[0050] This embodiment also provides a specific example:

[0051] According to the main factors affecting the gain of the fiber Raman amplifier, the value range of each influencing factor is determined. Among them, the wavelength range of the three pump lights is λ p1 =[1414 1437.3]nm, λ p2 =[1437.3 1460.6]nm, λ p3 =[1460.61484]nm, the power range of the pump light is P pi= [150 300] mW, where i = 1, 2, 3. Within this range, 6000 sets of pump power and wavelength data were randomly generated, and the corresponding fiber Raman amplifier gain values for each set of pump power and wavelength data were obtained. To verify the effectiveness of the proposed two-stage neural network-based adaptive gain control method for fiber Raman amplifiers, 5000 sets of data were used as training data, and the remaining 1000 sets of data were used as test data.

[0052] In order to speed up the convergence of the neural network, the training data is supplemented with missing data, processed for exceptions, and normalized. The expression for the normalization of the pump light wavelength value is: i =X i ÷70×0.99+0.01, where 70 represents the total wavelength range of the three pump lights. 0.01 is selected as the lowest point of normalization to avoid weight update failure caused by zero value input. The expression for the normalization of the pump light power value is: Y i =(X i -150)÷150×0.99+0.01; the expression for gain normalization is: Y i =(X i )÷X max ×0.99+0.01, where X max Represents the maximum gain value of the fiber Raman amplifier in the training data.

[0053] Determine the model structure of the primary neural network and the secondary neural network based on the training data, such as Figure 2 As shown in Figure 2, the input layer of the first-level neural network contains 40 neurons, G1(λ s1 )…G 40 (λ s40 ) represents the gain value of 40 different wavelength signal lights; the output layer of the first-level neural network contains 6 neurons, P p1 …P p3 Represents three pump light power values, λ p1 …λ p3 represents three pump light wavelength values; the number of hidden layers of the first-level neural network is 3 and each layer contains 100 neurons. The input layer of the second-level neural network is the output layer of the first-level neural network; the output layer of the second-level neural network contains 40 neurons, G1(λ s1 )…G 40 (λ s40 ) represents the gain value of 40 different wavelength signal lights; the number of hidden layers of the secondary neural network is 3 and each layer contains 100 neurons. The input layer, hidden layer and output layer of the reverse neural network and the forward neural network adopt the S-type transfer function sigmoid (y=1 / (1+e -x)) are connected in sequence. The initial connection weights of neurons in the neural network are set to random numbers in the range [-0.5 0.5].

[0054] The first-level neural network and the second-level neural network models are trained using the training data to obtain trained first-level neural network and second-level neural network.

[0055] Calculate the target gain value of the Raman amplifier using the trained first-level neural network model The pump light power and pump light wavelength under the first level neural network are calculated using the trained second level neural network model to calculate the fiber Raman amplifier gain value (y i (i=1…40)).

[0056] The gain value (y i (i=1…40)) Same as the target gain value of the fiber Raman amplifier Subtract the gain error vector error = G tar -G pre .

[0057] The weight values of each neuron link of the primary neural network and the secondary neural network are updated according to the gain error combined with the gradient descent method. The gradient descent method formula is: Where W s，t Represents the connection weight matrix of two adjacent layers, where the s layer is in front and the t layer is in the back. The error propagation formula of two adjacent layers is O s and O t Represent the output vectors of the s layer and t layer respectively.

[0058] The gain error between the predicted gain value and the target gain value is calculated again, and the neuron link weight value is adjusted cyclically using the gain error and the gradient descent method to finally obtain a highly accurate amplifier output gain value and the corresponding pump light power and wavelength values.

[0059] This algorithm model is implemented in the PyCharm environment, and the root mean square error can be used. Measure the accuracy of the fiber Raman amplifier gain spectrum before and after optimization. 1000 sets of test data were verified, and the results are as follows Figure 3 As shown in the figure, the mean of RMSE before optimization is 0.715 and the variance is 0.075; the mean of RMSE after optimization is 0.131 and the variance is 0.038.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fiber Raman amplifier gain adaptive control method based on a two-stage neural network, characterized in that: include: Acquiring training data, wherein the training data includes pump light wavelength data, pump light power data, and corresponding fiber Raman amplifier gain values; Determining the topological structures of the primary neural network and the secondary neural network based on the training data to obtain initial primary neural network and initial secondary neural network models; the primary neural network uses the fiber Raman amplifier gain value as an input layer and the pump light wavelength and pump light power as an output layer; the secondary neural network uses the pump light wavelength and pump light power as an input layer and the fiber Raman amplifier gain value as an output layer; Using the training data to train the primary neural network model and the secondary neural network model to obtain trained primary neural network model and secondary neural network model; Determining the topological structure of the primary neural network based on the training data to obtain an initial primary neural network model specifically includes: Determining an input layer of a primary neural network model according to the fiber Raman amplifier gain value in the training data; Determining a hidden layer of a primary neural network model according to a gain error of the optical fiber Raman amplifier; Determining an output layer of a primary neural network model according to wavelength data of the pump light and power data of the pump light in the training data; Determining the topological structure of the secondary neural network based on the training data to obtain an initial secondary neural network model specifically includes: Determining an input layer of a secondary neural network model according to wavelength data of the pump light and power data of the pump light in the training data; Determining a hidden layer of a secondary neural network model according to a gain error of the optical fiber Raman amplifier; Determining an output layer of a secondary neural network model according to the fiber Raman amplifier gain value in the training data; Calculating the pump light power and pump light wavelength of the fiber Raman amplifier at a target gain value according to the trained first-level neural network model; Calculating the predicted gain value of the fiber Raman amplifier at the pump light power and pump light wavelength output by the first-level neural network according to the trained second-level neural network model; Subtracting the predicted gain value obtained by the secondary neural network from the target gain value of the fiber Raman amplifier to obtain a gain error; updating the weight values of each neuron link of the primary neural network and the secondary neural network according to the gain error in combination with the gradient descent method, recalculating the gain error between the predicted gain value and the target gain value, and cyclically adjusting the neuron link weight values by using the gain error and the gradient descent method, ultimately obtaining a highly accurate amplifier output gain value and corresponding pump light power and wavelength values; The gradient descent method formula is: Among them, W s,t Represents the connection weight matrix of two adjacent layers, with the s layer in front and the t layer in the back; The error propagation formula between two adjacent layers is: O s and O t Represent the output vectors of the s layer and the t layer respectively.

2. The fiber Raman amplifier gain adaptive control method based on a two-stage neural network according to claim 1 is characterized in that Before determining the topological structures of the primary neural network and the secondary neural network models according to the training data, the method further includes: The training data is subjected to missing data completion, outlier processing, and normalization processing.

3. The fiber Raman amplifier gain adaptive control method based on a two-stage neural network according to claim 1, characterized in that: include: The input layer of the primary neural network contains 40 neurons, the number of hidden layers is 3 and each layer contains 100 neurons, and the output layer contains 6 neurons; The input layer of the secondary neural network includes 6 neurons, the number of hidden layers is 3 and each layer includes 100 neurons, and the output layer includes 40 neurons.

4. The fiber Raman amplifier gain adaptive control method based on a two-stage neural network according to claim 1, characterized in that: The input layer, hidden layer and output layer in the primary neural network and the secondary neural network are connected in sequence using an S-type transfer function sigmoid.

Images