光纤布拉格光栅
谱线
跟踪(教育)
计算机科学
校准
反射计
理论(学习稳定性)
约束(计算机辅助设计)
人工智能
培训(气象学)
光学
联轴节(管道)
材料科学
期限(时间)
稳健性(进化)
温度测量
直线(几何图形)
声学
算法
栅栏
卷积神经网络
模式识别(心理学)
控制理论(社会学)
时域
辐射
纤维
跟踪误差
工作(物理)
均方误差
作者
Bo Liu,Peidong Cao,Feng Xia,Yuting Cai,Rui Pei
标识
DOI:10.1088/1361-6501/aeb8a7
摘要
Abstract Locally overlapped spectra of identical weak-reflectivity fiber Bragg gratings in optical frequency domain reflectometry (OFDR) are susceptible to occasional large-error predictions when adjacent-peak coupling and target-peak degradation obscure shift-related spectral features. This work formulates robust shift tracking as a tail-error suppression problem and introduces Tail-aware training into a one-dimensional convolutional neural network. By combining a tail-penalty term with hard-sample weighting, the training objective increases the contribution of high-error samples and strengthens the constraint on difficult local spectra. An OFDR experimental system was established, and the method was evaluated under uniform-temperature-control baseline experiment, independent temperature control of adjacent gratings, and non-uniform heating of the target grating. For the non-uniform heating case, an operational equivalent reference wavelength shift was constructed from the average temperature at both ends of the target grating and the calibration relationship obtained under uniform heating. Experimental results show that the proposed CNN-Tail-aware (Tail-aware CNN)improves tracking stability and suppresses high-percentile errors under complex local spectral overlap. Based on the completed ten-seed means, its RMSE and P99 were reduced by 26.7% and 20.4%, respectively, compared with CNN-MAE (Baseline CNN). These results indicate that Tail-aware training provides a task-oriented way to reduce large-error prediction risk in robust tracking of locally overlapped spectra under OFDR conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI