基于神经网络的数字校准技术综述  

Review of Digital Calibration Techniques Based on Neural Network

作　　者：李嘉燊 李龙[1] 邓红辉[1,2] 陈红梅[1] 孟煦 尹勇生[1,2] LI Jiashen;LI Long;DENG Honghui;CHEN Hongmei;MENG Xu;YIN Yongsheng(Institute of VLSI Design,Hefei Univ.of Technol.,Hefei 230601,P.R.China;IC Design Web-Cooperation Research Center of MOE,Hefei Univ.of Technol.,Hefei 230009,P.R.China)

机构地区：[1]合肥工业大学微电子设计研究所,合肥230601 [2]合肥工业大学教育部IC设计网上合作研究中心,合肥230009

出　　处：《微电子学》2022年第2期191-196,共6页Microelectronics

基　　金：国家重点研发计划资助项目(2018YFB2202604);安徽省科技攻关计划资助项目(202104g01020008);安徽高校协同创新资助项目(GXXT-2019-030)。

摘　　要：随着集成电路工艺的发展以及晶体管尺寸的不断减小,ADC转换率变得更快、功耗更低,但器件的失配误差随之变得更大,从而影响精度,因此引入校准电路已成必然趋势。文章首先介绍了几种ADC的常见误差及其校准方法,然后介绍了神经网络的工作原理,并总结了几种主要的基于神经网络的数字校准方法,分析了不同方法的优势和劣势。最后,针对14位流水线ADC,给出了神经网络校准算法的系统级仿真验证结果。经校准后,有效位数(ENOB)从10位提升到12.5位,无杂散动态范围(SFDR)从80 dB提升到100 dB。With the development of integrated circuit process,the sizes of transistors are reducing,and the ADCs become faster with lower power consumption.On the other hand,smaller size brings more mismatch error,which will affect the accuracy,so it is necessary to introduce error calibration.In this paper,firstly,the typical error sources and traditional calibration methods of ADC were analyzed.Secondly,basic principle of neural network was given together with the most popular researches for ADC calibration based on neural network,and the advantages and disadvantages of different methods were analyzed.Finally,a system level simulation based on neural network was proposed to calibrate a 14-bit pipelined ADC.The results showed that ENOB was improved from 10 bit to 12.5 bit,and SFDR was improved from 80 dB to 100 dB.

关 键 词：模数转换器 全数字校准 神经网络 

分 类 号：TN792[电子电信—电路与系统]