基于深度学习的微电流测量  

MICRO-CURRENT MEASUREMENT BASED ON DEEP LEARNING

作　　者：刘欣 黄旭海 陇盛 蒋加平 LIU Xin;HUANG Xuhai;LONG Sheng;JIANG Jiaping(Army Artillery and Air Defense Academy,Hefei,Anhui 230031;National University of Defense Technology,Changsha,Hunan 410073;Anhui Changjiang Institute of Metrology&Measurement,Hefei,Anhui 230088)

机构地区：[1]陆军炮兵防空兵学院基础部,安徽合肥230031 [2]国防科技大学信息系统工程重点实验室,湖南长沙410073 [3]安徽省长江计量所,安徽合肥230088

出　　处：《物理与工程》2022年第3期115-120,共6页Physics and Engineering

基　　金：国家自然科学基金资助项目(62076252)。

摘　　要：在微电流测量中,抑制噪声干扰是重点问题。针对目前消除噪声干扰多集中在装置设计阶段,提出一种借助深度学习技术在数据处理阶段消除噪声干扰的测量方法。以微电流测量装置的输出电压值和计量所高精度万用表的微电流相对真值构成数据样本,选用Tensorflow实验平台,利用1500个数据样本训练“最优神经网络”以对微电流值进行预测,并利用剩余的27个样本检验。实验结果表明,较传统最小二乘法解得的“最佳函数”,深度学习方法能较好消除噪声干扰,平均相对误差更小、预测精度更高。In micro-current measurement, the suppression of noise interference is a key issue. Aiming at the current elimination of noise interference, which is mostly concentrated in the device design stage, a measurement method is proposed to eliminate noise interference in the data processing stage with the aid of a deep learning technology. The output voltage value of the micro-current measuring device and the relative true value of the micro-current of the high-precision multimeter of the measurement are used to form data samples. The Tensorflow experimental platform is selected and the “optimal neural network” is trained using 1500 data samples to predict the micro-current value, and the remaining 27 samples are used to test. The experimental results show that, compared with the “best function” obtained by the traditional least square method, the deep learning method can better eliminate noise interference, with smaller average relative errors and higher prediction accuracy.

关 键 词：微电流 深度学习 神经网络 最小二乘法 

分 类 号：TM933.1[电气工程—电力电子与电力传动] TP18[自动化与计算机技术—控制理论与控制工程]