基于神经网络和多目标优化算法的掺铋光纤放大器设计
Multi-band transmission is considered an effective solution to address the increasing capacity constraints in fiber optic communication systems. However, due to the lack of mature optical amplifiers, the large-scale deployment of dense wavelength division multiplexing (DWDM) technology for long-distance transmission in bands such as O, E, and S has not yet been achieved. In recent years, researchers have discovered that different dopants in bismuth-doped silica fibers exhibit broad fluorescence characteristics in the near-infrared region. This finding brings hope for addressing the aforementioned challenges. In traditional approaches, the performance analysis of amplifiers often requires solving a set of coupled differential equations using methods such as the Runge-Kutta algorithm combined with the Shooting method or Relaxation method. When incorporating global optimization algorithms, it becomes necessary to solve thousands of related equations, resulting in a complex and time-consuming process. Previous research methods have mainly focused on the optimization design of Raman fiber amplifiers or hybrid optical amplifiers, with fewer studies specifically targeting the structural optimization design of doped fiber amplifiers, particularly bismuth-doped fiber amplifier (BDFA). Moreover, most of these studies have employed single-objective optimization algorithms, resulting in obtaining only one optimal solution at a time. In general, there is a trade-off relationship between the gain and noise performance of amplifiers. Increasing the gain often leads to the deterioration of the noise performance, and vice versa. As a result, there is no unique optimal solution. Therefore, it is necessary to design a method that can accurately model the amplifier and efficiently optimize multiple performance metrics simultaneously.
The backpropagation neural network (BPNN) is a type of multilayer feedforward neural network consisting of input layer, hidden layers, and output layer. The input layer contains six neurons corresponding to the input signal wavelength and five structural parameters of the amplifier. The output layer contains two neurons corresponding to the Gain and noise figure (NF) of the respective wavelength signals. The main characteristic of BPNN is the forward propagation of signals and the backward propagation of errors. It belongs to the supervised learning methods. For multi-objective problems, the objective values are typically mutually constrained, and there is no unique optimal solution. Using multi-objective optimization algorithms can provide a set of independent optimal solutions, allowing engineering designers to choose based on their actual needs. NSGA-II is a multi-objective optimization algorithm that improves upon the non-dominated sorting genetic algorithm (NSGA). By introducing fast non-dominated sorting, elite preservation strategy, and crowding distance operator, NSGA-II reduces computational complexity, improves optimization efficiency, and ensures the diversity of individuals in the population.
Simulation experiments were conducted using a theoretical model of a two-stage BDFA to obtain a sample set. The BPNN model was trained and tested with different sample sizes, with a training-to-testing set ratio of 9∶1. It was observed that as the sample set size increased, the overall trend of RMSE decreased while the R2 value increased (Fig.4). When the sample size reached 3000, the BPNN model achieved an RMSE of 0.191 for Gain and 0.084 for NF in the testing phase, with R2 values of 0.999 and 0.998, respectively. The established BPNN model exhibits high prediction accuracy and can effectively capture the nonlinear relationship between the structural parameters and performance of the two-stage BDFA. Based on the established BPNN model, the objective function is evaluated, and after 100 iterations, a Pareto optimal solution set containing 500 solutions is obtained (Fig.6). Furthermore, a comparison is made between the performance of using SVM and BPNN for predicting Gain and NF. The results show that the BPNN model has smaller prediction errors and higher accuracy in predicting Gain and NF. Additionally, the time required for optimization design using BPNN-NSGA-II is five orders of magnitude lower than using Relaxation method combined with NSGA-II, taking less than 80 seconds to complete the design. Compared to SVM-NSGA-II, the time is reduced by one order of magnitude (Fig.9).
This paper proposes a multi-objective optimization method that combines BPNN and NSGA-II algorithms for accurate modeling and efficient design optimization of two-stage BDFA. By establishing a BPNN model to map the nonlinear relationship between structural parameters and performance, it avoids the need for repetitive solving of coupled differential equations. After training and testing, the BPNN model exhibits low RMSE and high R2 values. Using this BPNN model in conjunction with the NSGA-II algorithm, a Pareto optimal solution set containing 500 solutions is obtained. The paper also provides the Gain and NF spectra for five different amplifier configurations. Compared to other methods, the proposed approach significantly reduces the optimization design time, improves optimization efficiency, and enables the simultaneous attainment of multiple optimal solutions, providing decision-makers with more choices.
1 引言
多波段传输被认为是解决光纤通信系统容量日益紧缺的一种有效方法[1]。G652.D单模光纤在除C、L波段外的O、E、S等波段的传输损耗均低于0.4 dB/km。然而,由于缺少成熟的光放大器,至今尚未在这些波段大规模部署密集波分复用技术进行远距离传输[2-3]。近年来,随着掺铋光纤放大器(BDFA)的提出和发展[4-7],研究人员发现不同组分的掺铋石英光纤在近红外区域具有广泛的荧光特性[8-10],这一发现为解决上述难题带来了希望。就光纤放大器而言,通常以高增益、低噪声系数为目标,对结构参数进行优化设计。传统情况下,需要利用Runge-Kutta算法结合打靶法或松弛法求解耦合微分方程组以分析放大器的性能[11-13]。若结合全局优化算法,例如遗传算法、粒子群算法等来优化光纤放大器结构,则需要进行成千上万次相关方程组的求解,过程复杂且耗时。而通过机器学习的方法建立光纤放大器结构参数和性能之间的非线性映射关系,可以代替传统求解微分方程组的过程,大幅降低计算时间,从而实现对放大器结构的高效优化设计。
2015年,Singh等[14]使用遗传算法对EDFA拉曼混合光放大器的结构参数进行优化。同年,陈静等[15]使用最小二乘支持向量回归机建立多泵浦拉曼光纤放大器模型,并结合遗传算法实现在设计时间不超过19 s的情况下,快速调节开关增益。2018年,Chen等[16]使用极限学习机和差分进化算法相结合的混合优化算法对多泵浦拉曼光纤放大器进行优化设计,该算法缩短了计算时间,提高了整体优化效率。同年,Chen等[17]还使用最小二乘支持向量机结合粒子群算法来提高拉曼光纤放大器的优化效率。2021年,巩稼民等[18]采用神经网络建立泵浦波长和泵浦功率与拉曼净增益谱之间的映射关系,并结合人工蜂群算法来优化泵浦光参数。由此可见,上述方法主要对拉曼光纤放大器或者混合光放大器进行优化设计,少有用于对掺杂光纤放大器的结构优化设计。且上述方法大多使用单目标优化算法,每次优化只能得到一个最优解。而通常放大器的多个性能指标相互制约,不存在唯一的最优解,因此使用多目标优化算法进行优化设计是一个更好的选择。
本文采用带精英保留策略的快速非支配排序遗传算法(NSGA-Ⅱ)优化设计BDFA结构。该多目标优化算法可将多个波长信号的增益和噪声系数作为目标,生成Pareto最优解集。在评价个体适应度方面,通过建立反向传播神经网络(BPNN)模型代替求解耦合微分方程组,提升整体优化效率。最后,通过分析各个结构参数及对应放大器输出性能,得到两级分段泵浦式放大方案,为BDFA的结构设计提供指导。
2 相关理论分析
2.1 掺铋光纤放大器
光纤放大器的主要性能参数包括:信号增益(Gain)、噪声系数(NF),以及增益带宽等。其中,Gain和NF是优化BDFA所考虑的主要目标。为了获得这两个参数,通常需要利用Runge-Kutta算法结合打靶法或者松弛法对功率传播方程和粒子速率方程进行求解[19]。当将光纤放大器用于线路放大时,需要同时满足高增益和低噪声系数这两个基本条件[20]。然而,使用单级光纤进行放大很难同时满足这两个条件。为实现高增益,需要使用较长的掺杂光纤,但这会造成自发辐射噪声的累积,从而使噪声系数上升。另一方面,为实现低噪声系数,需要使用较短的光纤来保证较高的总体反转率,但这样又无法获得理想的增益。在以往的光纤放大器结构中,通常会采用多级放大的方式来改善线路放大器的性能。在多级放大方式下,整个放大器的噪声系数很大程度上取决于第一级放大器的噪声系数。因此,若将第一级设计为低噪声放大器,以保证较好的噪声性能,将后续级设计为高效的功率放大器,以保证较好的增益性能,这样可以使得整个放大器同时具备高增益和低噪声系数[21]。
本研究采用掺铋磷硅酸盐光纤作为增益介质,该光纤在1240 nm泵浦光的激发下,能够实现对O波段附近信号光的放大[8]。掺铋光纤(BDF)中铋离子浓度较低,因此需要使用较长的光纤来实现所需的信号放大,长度通常为100~200 m。放大器采用两级分段泵浦式结构,如
对于两级BDFA,每种结构参数配置下需要进行两次耦合微分方程组的求解操作,即以第一级光纤的末端值和第二级的正向泵浦功率作为第二级的初始值,然后进行第二次求解。整个过程相对复杂且非常耗时,平均每次计算需要50 s。在某些特殊情况下,还可能无法得到耦合微分方程组的收敛解。如果采用智能优化算法,例如遗传算法、粒子群算法等对放大器结构参数进行优化,每次迭代优化都需要进行成千上万次耦合微分方程组的求解操作,将花费更长的计算时间。
2.2 反向传播神经网络
BPNN是一种多层前馈神经网络模型[23],具有较强的非线性映射能力和自适应能力。由于其原理和结构简单,对于小数据集模型,更容易避免过拟合问题。因此本研究利用其来映射两级BDFA的结构参数和性能之间的非线性关系。BPNN主要包括输入层、隐藏层和输出层,如
式中:H是隐藏层神经元数量;I是输入层神经元数量;O是输出层神经元数量;a是介于1~10之间的常数。各层之间通过激活函数联系起来,激活函数的主要作用是引入非线性因素,以增强神经网络的表达能力。
BPNN的主要特点是信号的前向传递和误差的反向传播。它属于有监督学习方法,需要使用一组已知目标输出的学习样本集进行训练。训练过程中,首先使用随机值初始化网络的权值和阈值。然后,输入信号从输入层开始,经过隐藏层逐层处理并向前传递,直到输出层。如果输出层的预测值与期望输出不符,则进行误差的反向传播。利用梯度下降法,逐层调整神经网络的权值和阈值。这个调整的过程是反复进行的,直到网络的预测值逼近目标值且误差不再下降时,训练完成。
2.3 多目标优化算法
本研究涉及的多目标问题可由
式中:f1(
NSGA-Ⅱ是在非支配排序遗传算法(NSGA)基础上改进的一种多目标优化算法[24]。通过引入快速非支配排序方法、精英保留策略和拥挤度算子,降低了计算复杂度、提高了优化效率并能保证种群中个体的多样性。算法流程如
1)初始化迭代次数
2)对种群Pt执行交叉和变异操作,生成大小为N的种群Qt。合并种群Pt和Qt,生成大小为2N的种群Rt。
3)使用经过训练的BPNN模型对种群Rt中每个个体进行评估,即获得每种参数配置下的两级BDFA性能,包括信号增益和信号噪声系数。
4)根据评估得到的适应度,对种群Rt中的个体执行快速非支配排序操作,确定每个个体的Pareto等级。从Pareto第一等级开始,将其个体加入下一代种群Pt+1中,直到Pt+1的大小为N。
5)若Pareto第i等级加入后,种群Pt+1大小超过N,而第i-1等级加入后,Pt+1大小小于N,则计算第i等级中每个个体沿着目标两侧最近的两个相邻个体之间的平均距离。该距离称为拥挤距离,选择拥挤距离更大的个体进入下一代,直到种群Pt+1的大小为N。
6)重复步骤2)~5),直到达到最大迭代次数时停止。此时的P为Pareto最优解集,即具有最优增益性能和噪声性能的一系列放大器结构参数配置。
3 仿真与分析
3.1 BPNN模型的建立
利用两级BDFA的理论模型进行仿真实验,通过求解功率传播方程以获得样本集。放大器的结构参数包括两级BDF的长度、两级泵浦功率和第二级前后向泵浦功率比。除结构参数外,还将输入信号波长作为仿真的变量参数,而输入信号功率统一设置为-20 dBm。相应参数的取值范围见
表 1. 放大器仿真变量参数取值范围
Table 1. The ranges of parameters in the simulation
|
通过试错法,综合评估训练时间和预测精度后,最终确定神经网络的结构为6-12-8-2。其中,输入层和隐藏层、隐藏层和隐藏层之间的激活函数选用tansig函数,而隐藏层和输出层之间的激活函数选用purelin函数。采用Levenberg-Marquardt(L-M)方法训练BPNN,该方法的收敛速度较快并且均方误差较小,适合训练中小规模的神经网络。具体的训练参数设置如下:学习率为0.0001,最大迭代次数为1000,训练目标的最小误差为0.001。模型的准确度由均方根误差(ERMSE)和决定系数(R2)来衡量:
式中:
分别在不同样本数量下,训练BPNN并进行测试,训练集和测试集大小的比值为9∶1。
图 4. 不同样本数量训练BPNN模型,对Gain和NF的测试性能
Fig. 4. The test performance of Gain and NF after training BPNN model with different sample sizes
接下来,随机选取2700个样本作为训练集,300个样本作为测试集,对BPNN模型进行训练和测试。
图 5. 预测值与目标值的拟合情况。(a)(b)BPNN预测的Gain和NF
Fig. 5. Fitting of predicted value and target value. (a)(b) Gain and NF predicted by BPNN
3.2 多目标优化
使用NSGA-Ⅱ算法进行优化设计。将放大器的结构参数作为个体特征,初始化种群。根据
为了方便表示,从Pareto最优解集中选取5组具有代表性的解,如
表 2. 部分Pareto最优解集及相应的适应度值
Table 2. Partial Pareto optimal solution sets and corresponding fitness values
|
3.3 性能比较
为进行比较,同时建立支持向量机(SVM)模型用于映射两级BDFA的结构参数和性能之间的非线性关系。选用径向基函数作为核函数,惩罚因子和gamma参数分别设置为20和0.9。在样本数量同样为3000、训练集和测试集按9∶1的比例划分下,对SVM模型进行训练并测试。
图 8. 使用BPNN模型和SVM模型对Gain和NF测试的绝对误差PDF曲线。(a)Gain;(b)NF
Fig. 8. PDF curves of testing absolute errors for Gain and NF using BPNN and SVM models. (a) Gain; (b) NF
最后,记录使用松弛法求解一次两级BDFA耦合微分方程组花费的时间,若与NSGA-Ⅱ算法结合进行优化设计,预测将花费106 s。同时分别记录使用SVM和BPNN与NSGA-Ⅱ算法结合进行优化所需花费的时间,如
4 结论
提出一种结合神经网络和NSGA-II算法的多目标优化方法,用于准确建模和高效优化设计两级掺铋光纤放大器。通过建立BPNN模型映射结构参数和性能之间的非线性关系,避免了耦合微分方程组的大量重复求解过程。经过训练和测试,BPNN模型表现出较低的ERMSE和较高的R2值。利用该BPNN模型结合NSGA-II算法,得到了一个包含500个解的Pareto最优解集,并给出5组不同放大器结构参数配置下的信号增益和噪声系数谱。与其他方法的比较结果表明,所提方法显著降低了优化设计时间,提升了优化效率,并能同时得到多个最优解,为决策者提供更多选择。综上所述,所提方法对掺铋光纤放大器以及其他掺杂光纤放大器的高效率、多目标结构优化设计具有重要意义。
[1] Deng N, Zong L J, Jiang H Y, et al. Challenges and enabling technologies for multi-band WDM optical networks[J]. Journal of Lightwave Technology, 2022, 40(11): 3385-3394.
[2] Renaudier J, Napoli A, Ionescu M, et al. Devices and fibers for ultrawideband optical communications[J]. Proceedings of the IEEE, 2022, 110(11): 1742-1759.
[3] Rapp L, Eiselt M. Optical amplifiers for multi-band optical transmission systems[J]. Journal of Lightwave Technology, 2022, 40(6): 1579-1589.
[4] Khegai A, Ososkov Y, Firstov S, et al. Gain clamped Bi-doped fiber amplifier with 150 nm bandwidth for O- and E-bands[J]. Journal of Lightwave Technology, 2022, 40(4): 1161-1166.
[5] Dianov E M. Amplification in extended transmission bands using bismuth-doped optical fibers[J]. Journal of Lightwave Technology, 2013, 31(4): 681-688.
[6] Bufetov I A, Melkumov M A, Firstov S V, et al. Bi-doped optical fibers and fiber lasers[J]. IEEE Journal of Selected Topics in Quantum Electronics, 2014, 20(5): 111-125.
[7] KhegaiA, OsoskovY, FirstovS, et al. O+E band BDFA with flattop 116 nm gain bandwidth pumped with 250 mW at 1256 nm[C]//Optical Fiber Communication Conference (OFC) 2021, June 6-10, 2021, Washington, DC. Washington, D.C.: Optica Publishing Group, 2021: Tu1E.4.
[8] Firstov S V, Khegai A M, Kharakhordin A V, et al. Compact and efficient O-band bismuth-doped phosphosilicate fiber amplifier for fiber-optic communications[J]. Scientific Reports, 2020, 10: 11347.
[9] Firstov S V, Khopin V F, Bufetov I A, et al. Combined excitation-emission spectroscopy of bismuth active centers in optical fibers[J]. Optics Express, 2011, 19(20): 19551-19561.
[10] Firstova E G, Bufetov I A, Khopin V F, et al. Luminescence properties of IR-emitting bismuth centres in SiO2-based glasses in the UV to near-IR spectral region[J]. Quantum Electronics, 2015, 45(1): 59-65.
[11] Shang L, Qi L M, Liao Y L, et al. A combined algorithm for simulating fiber lasers based on the shooting and relaxation methods[J]. Optical Fiber Technology, 2012, 18(6): 502-508.
[12] Lali-Dastjerdi Z, Kroushavi F, Rahmani M H. An efficient shooting method for fiber amplifiers and lasers[J]. Optics & Laser Technology, 2008, 40(8): 1041-1046.
[13] Liu X M. Fast methods for solving nonlinearly coupled equations describing fiber amplifiers[J]. Applied Mathematics and Computation, 2006, 172(2): 1018-1035.
[14] Singh S, Kaler R S. Performance optimization of EDFA-Raman hybrid optical amplifier using genetic algorithm[J]. Optics & Laser Technology, 2015, 68: 89-95.
[15] 陈静, 江灏, 刘暾东, 等. 基于最小二乘支持向量回归模型的拉曼光纤放大器优化设计[J]. 光学学报, 2015, 35(11): 1123004.
[16] Chen J, Jiang H. Optimal design of gain-flattened Raman fiber amplifiers using a hybrid approach combining randomized neural networks and differential evolution algorithm[J]. IEEE Photonics Journal, 2018, 10(2): 7101915.
[17] Chen J, Qiu X J, Yin C Y, et al. Efficient design of gain-flattened multi-pump Raman fiber amplifiers using least squares support vector regression[J]. Journal of Optics, 2018, 20(2): 025702.
[18] 巩稼民, 刘芳, 吴艺杰, 等. 基于神经元网络和人工蜂群算法的拉曼光纤放大器设计方案[J]. 光学学报, 2021, 41(20): 2006002.
[19] Giles C R, Desurvire E. Modeling erbium-doped fiber amplifiers[J]. Journal of Lightwave Technology, 1991, 9(2): 271-283.
[20] Delavaux J M P, Nagel J A. Multi-stage erbium-doped fiber amplifier designs[J]. Journal of Lightwave Technology, 1995, 13(5): 703-720.
[21] Mishra A R, Kakade R N, Kakade P D. Dual-stage EDFA for improving the performance of long-haul optical systems[J]. IEEE Access, 2022, 10: 13496-13514.
[22] Wang D C, Pei L, Zheng J J, et al. Analysis of gain and noise characteristics of O-band Bi-doped fiber amplifier under different pumping schemes[J]. Optik, 2022, 251: 168491.
[23] Hannan M A, Lipu M S H, Hussain A, et al. Neural network approach for estimating state of charge of lithium-ion battery using backtracking search algorithm[J]. IEEE Access, 2018, 6: 10069-10079.
[24] Deb K, Pratap A, Agarwal S, et al. A fast and elitist multiobjective genetic algorithm: NSGA-II[J]. IEEE Transactions on Evolutionary Computation, 2002, 6(2): 182-197.
[25] Zhang Y P, Liu D, Shen X E, et al. Design of iodine absorption cell for high-spectral-resolution lidar[J]. Optics Express, 2017, 25(14): 15913-15926.
[26] Verma S, Pant M, Snasel V. A comprehensive review on NSGA-II for multi-objective combinatorial optimization problems[J]. IEEE Access, 2021, 9: 57757-57791.
[27] 江灏, 林朝晖, 陈静, 等. 考虑非对称重叠光谱的FBG传感网络带宽资源多目标优化配置[J]. 光学学报, 2021, 41(10): 1006002.
[28] 王燕燕, 李家豪, 舒林森, 等. 基于RSM与NSGA-Ⅱ算法的激光熔覆参数多目标优化[J]. 激光与光电子学进展, 2022, 59(7): 0714004.
Article Outline
侯文强, 裴丽, 王建帅, 郑晶晶, 徐文轩, 田梓辰, 王丁辰, 王丽红, 李晶, 宁提纲. 基于神经网络和多目标优化算法的掺铋光纤放大器设计[J]. 光学学报, 2024, 44(2): 0206001. Wenqiang Hou, Li Pei, Jianshuai Wang, Jingjing Zheng, Wenxuan Xu, Zichen Tian, Dingchen Wang, Lihong Wang, jing Li, Tigang Ning. Design of Bismuth-Doped Fiber Amplifier Based on Neural Network and Multi-Objective Optimization Algorithm[J]. Acta Optica Sinica, 2024, 44(2): 0206001.