基于神经网络的光纤温度估算方法的优化  

Optimized Neural Network Method for Temperature Estimation along Optical Fiber

作　　者：李苏雅[1] 董艳唯[1] 李琳[1] 张弛[1] 李楠 宁琦 陈永辉 LI Suya;DONG Yanwei;LI Lin;ZHANG Chi;LI Nan;NING Qi;CHEN Yonghui(Electric Power Science Research Institude,State Grid Tianjin Electric Power Company,Tianjin 300384,China;Department of Electronic and Communication Engineering,No)

机构地区：[1]国网天津市电力公司电力科学研究院,天津300384 [2]华北电力大学电子与通信工程系,河北保定071003

出　　处：《光通信研究》2025年第1期83-88,共6页Study on Optical Communications

基　　金：国家自然科学基金资助项目(62171185)。

摘　　要：【目的】为了有效估算基于布里渊散射的分布式光纤传感中光纤的温度,文章将多层前馈人工神经网络(ANN)应用于温度的估算。【方法】文章在Matlab软件中编写了用于光纤温度计算的单斜坡法、基于伪Voigt模型的最小二乘拟合法和ANN程序,同时仿真产生了不同信噪比(SNR)下的布里渊谱,采用以上3种算法计算了光纤温度,验证了ANN方法的有效性。在此基础上基于以上仿真产生的布里渊谱研究了ANN的关键参数,即隐层数量、隐层神经元数量和训练目标对训练速度、温度计算时间和准确性的影响规律。【结果】结果表明,ANN方法在SNR为22和37 dB时最大温度误差分别仅为1.18和0.63℃,且计算时间仅为最小二乘拟合法的1/1000左右。当隐层神经元数量不变时,随着隐层层数的增加,训练时间明显下降,计算时间线性增加,但其对温度估算的准确性几乎无影响;随着隐层神经元数量的增加,训练时间和计算时间均增加,隐层有21个神经元时,训练时间近似为1个神经元的67倍,但其对温度估算的准确性几乎无影响;训练目标(布里渊频移误差的平方)小于临界值(约为1 MHz 2)时,随着训练目标的增加,温度误差几乎不变,超过临界值后,随着训练目标的增加,温度误差增大。【结论】采用多层前馈ANN应用于基于布里渊散射的分布式光纤传感中的光纤温度估算时,建议选择单隐层且隐层神经元选择1个,训练目标选择1 MHz 2。【Objective】To effectively estimate the optical fiber temperature of distributed optical fiber sensing based on Brillouin scattering,a multi-layer feedforward Artificial Neural Network(ANN)is introduced to estimate the temperature.【Methods】The article has developed programs in Matlab for fiber optic temperature calculation using the single slope method,the least squares fitting method based on the pseudo-Voigt model,and an ANN.At the same time,the Brillouin spectra with different Signal-to-Noise Ratios(SNR)are simulated.Based on the Brillouin spectrum generated from the above simulations,the study investigates the key parameters of the ANN,namely the number of hidden layers,the number of neurons in the hidden layers,and the training objectives,and their influence on training speed,temperature calculation time,and accuracy.【Results】The maximum temperature error of ANN is only 1.18 and 0.63℃at 22 and 37 dB respectively,and the calculation time of ANN is only about 1/1000 of that of the least-squares fit method.When the number of hidden layer neurons remains constant,the training time decreases obviously with the number of hidden layer and the computation time increases linearly with the number of hidden layer.However,it has little effect on the accuracy of temperature estimation.Both the training time and computation time increase with the number of neurons in the hidden layer.When there are 21 neurons in the hidden layer,the training time is approximately 67 times that of only one neuron in the hidden layer.However,it also has little effect on the accuracy of temperature estimation.When the training goal(square of the Brillouin shift error)is less than the critical value(about 1 MHz 2),the temperature error is almost independent.However,when the training goal exceeds the critical value,the temperature error increases with the training goal.【Conclusion】When ANN is used to estimate optical fiber temperature of distributed optical fiber sensing based on Brillouin scattering,it is recommended to selec

关 键 词：分布式光纤传感 布里渊散射 布里渊频移 人工神经网络 温度 优化 

分 类 号：TN247[电子电信—物理电子学]