基于GRU神经网络的恒温晶振频率漂移预测  

Frequency Offset Prediction of Oven Controlled Crystal Oscillator Based on GRU Neural Network

作　　者：王彬 晏学成 张海利 刘岳巍 WANG Bin;YAN Xuecheng;ZHANG Haili;LIU Yuewei(School of Electrical and Electronic Engineering,Shijiazhuang Tiedao University,Shijiazhuang Hebei 050043,China;Hebei Far East Communication System Engineering Co.,Ltd.,Shijiazhuang Hebei 050200,China)

机构地区：[1]石家庄铁道大学电气与电子工程学院,河北石家庄050043 [2]河北远东通信系统工程有限公司,河北石家庄050200

出　　处：《电子器件》2024年第2期502-508,共7页Chinese Journal of Electron Devices

摘　　要：恒温晶振作为新时代5G通信系统的重要器件,其频率稳定性相当重要,准确预测恒温晶振频率漂移可以提高系统工作状态的安全性与可靠性。为进一步提高频率漂移预测精度,满足5G通信系统需求,提出了一种基于门控循环单元(GRU)神经网络的晶振频率预测模型。该模型同时考虑温度、老化两种因素,并利用神经网络具有出色的自适应性与非线性泛化能力,学习恒温晶振频率漂移变化规律。最后以14 d实测数据为例进行研究分析,并与循环神经网络、长短时记忆神经网络做对比实验,以均方根误差、平均绝对误差、算法运行时间作为评价指标,结果表明GRU网络具有更高的预测精度、更快的运算速度。The frequency stability of oven controlled crystal oscillator,an important component of 5G communication system in the new era is very important.Accurate prediction of oven controlled crystal oscillator frequency offset can improve the security and reliability of system working state.In order to further improve the prediction accuracy of frequency offset and meet the requirements of 5G communi-cation system,a crystal frequency prediction model based on gated recurrent unit(GRU)neural network is proposed.The model takes into account both temperature and aging factors,and uses neural network with excellent adaptability and nonlinear generalization ability to learn the change rule of oven controlled crystal oscillator frequency offset.Finally,the measured data of 14 days are analyzed and compared with those got by using recurrent neural network and long short-term memory neural network.Root mean square error,mean absolute error and algorithm running time are taken as evaluation indexes,the results show that GRU network has higher prediction ac-curacy and faster operation speed.

关 键 词：恒温晶振 频率漂移预测 温度 老化 GRU神经网络 

分 类 号：TN752[电子电信—电路与系统]