| 基于卷积神经网络的混叠光谱解调方法 |
| |
| 引用本文： | 刘瀚霖,辛璟焘,庄炜,夏嘉斌,祝连庆.基于卷积神经网络的混叠光谱解调方法[J].红外与激光工程,2022,51(5):20210419-1-20210419-9. |
| |
| 作者姓名： | 刘瀚霖 辛璟焘 庄炜 夏嘉斌 祝连庆 |
| |
| 作者单位： | 1.北京信息科技大学 光电测试技术与仪器教育部重点实验室，北京 100192 |
| |
| 基金项目： | 国家自然科学基金(51535002，61903042)；北京市自然科学基金(4212048)；高等学校学科创新引智计划(D17021)；北京信息科技大学师资补充与支持计划(5029011103) |
| |
| 摘 要： | 研究了一种基于深度学习的光纤光栅混叠FBG光谱解调方法。利用卷积神经网络(Convolutional Neural Networks, CNN)模型处理混叠光谱非线性序列模型问题，通过一维卷积神经网络预测识别混叠光谱中心波长，并搭建了并联结构的混叠光谱数据自动采集实验系统，验证了混叠光谱的中心波长高精度解调。实验分析了训练样本、迭代次数对训练时间、测试时间、解调精度的影响，并对训练完成后的模型进行了解调时间测试。分别与其他解调算法进行了解调精度和测试时间对比，同时对同一组光谱数据使用解调模型算法及最高点寻峰算法进行中心波长值的对比并进行误差分析。实验结果表明:解调模型均方根误差结果为0.082 58 pm，使用Intel(R) Core(TM) i7-8550U CPU （Central Processing Unit）的解调计算时间为0.338 s。研究结果表明:采用卷积神经网络模型对于混叠光谱中心波长解调结果的准确性具有合理性，与其他算法相比，文中的解调算法在解调精度和时间上具有优势，模型大小在400 kB以下，所需算力较小，可部署在小型嵌入式设备中，在大规模机载传感网络，结构健康监测中有良好的应用前景。
|
| 关 键 词： | 光纤光栅 混叠光谱 光谱数据采集系统 卷积神经网络 |
| 收稿时间： | 2021-12-25 |
Demodulation method of overlapping spectrum based on convolutional neural network |
| |
| Affiliation: | 1.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument, Beijing University of Information Technology, Beijing 100192, China2.Beijing Laboratory of Optical Fiber Sensing and Systems, Beijing Information Science & Technology University, Beijing 100016, China3.School of Instrument Science and Opto-electronics Engineering, Hefei University of Technology, Hefei 230009, China |
| |
| Abstract: | An FBG spectral demodulation method based on deep learning was studied. The Convolutional Neural Networks(CNN) model was used to deal with the nonlinear sequence model of the overlapping spectrum, and the central wavelength of the overlapping spectrum was predicted and identified through a one-dimensional convolutional neural network. And a parallel structure of the overlapping spectrum data automatic acquisition experimental system was built to verify the high-precision demodulation of the center wavelength of the overlapping spectrum. The experiment analyzes the effects of training samples and epoch times on training time, testing time, and demodulation accuracy, and tests the computational demodulation time of the model after training. The demodulation accuracy and test time were compared with other demodulation algorithms. At the same time, the demodulation model algorithm and the peak finding algorithm at the highest point were used to compare the center wavelength value and analyze the error for the same set of spectral data. The experimental results show that the root means square error of the demodulation model is 0.082 58 pm, and the demodulation calculation time is 30.886 ms, which is used Intel(R) Core (TM) i7-8550U CPU. The research results show that the convolutional neural network model is reasonable for the accuracy of the central wavelength demodulation results of the overlapping spectrum. Compared with other algorithms, the demodulation algorithm in this article has advantages in demodulation accuracy and time. The model size is less than 400 kB, and the required computing power is small. It can be deployed in small embedded devices. It has good application prospects in large-scale airborne sensor networks and structural health monitoring. |
| |
| Keywords: | |
|
| 点击此处可从《红外与激光工程》浏览原始摘要信息 |
|
点击此处可从《红外与激光工程》下载全文 |
|
