【作者】 刘洋；

【导师】 张万里；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 随着智能物联网、人工智能及大数据分析技术的快速发展,全球数据生成量呈现指数级增长态势。在传统数据处理架构中,基于冯诺依曼结构的集中式数据处理模式要求感知终端将采集数据全部上传至数据中心进行处理,面临着“功耗墙”问题和“存储墙”问题。在此背景下,边缘计算受到广泛关注,其中基于非易失性存储特性的忆阻器凭借其独特的并行计算能力、低功耗特性及高集成度优势,被视为突破边缘计算瓶颈的关键器件。相较于传统传感器,热流传感器在高速响应以及激光激励下直接生成电压脉冲方面展现出显著优势。但在基于忆阻器实现热流传感信号神经网络边缘计算的过程中面临着忆阻器非理想特性、输出电压脉冲幅值不匹配以及复杂网络拓扑引发的外围电路设计等问题,导致基于忆阻器实现热流传感信号处理的研究近乎空白。为此,本研究通过构建二值神经网络,系统表征了LiNbO₃单晶忆阻器的非理想电学特性（高阻态阻值波动、低阻态电导衰减）,基于电导-权重映射模型定量解析器件特性对网络识别精度的影响,最终验证了二值神经网络对忆阻器非理想特性的强鲁棒性。在此基础上,本研究构建了基于忆阻器的传感器端二值神经网络边缘计算验证平台。通过设计模块化外围电路将忆阻器与热流传感器进行耦合,通过激光轨迹的实时分类场景进行功能验证。实验选用z,v,n三个字母的标准字母形状的激光轨迹进行原理性验证,基于忆阻器实现了对热流传感信号的二值神经网络边缘计算,实现了对激光轨迹的分类识别。主要研究工作如下:（1）开展了忆阻器非理想特性对二值化神经网络识别准确率的影响研究。基于单晶LiNbO₃薄膜忆阻器对二值神经网络的忆阻特性进行研究,搭建了9×3规模的二值神经网络并通过3×3像素规模的z,v,n字母数据集对其进行训练得到二值神经网络权重矩阵。着重地对忆阻器的数据保持特性进行研究,通过python语言进行建模将忆阻器考虑低阻态衰减和高阻态波动的非理想特性映射到二值神经网络权重矩阵中当中,通过标准字母和标准字母的变形输入测试和分析非理想特性对二值神经网络识别准确率的影响。（2）开展了基于LiNbO₃忆阻器的热流传感信号二值神经网络边缘计算研究。对基于热流传感器的二值神经网络边缘计算系统基本原理与设计方法进行研究,根据热流传感器在激光照射下产生电势差的原理将激光轨迹转化成传感信号,并将忆阻器的高低两个阻态与二值神经网络中0,1两个权重值进行对应,随后开展了热流传感器与忆阻器的耦合方式研究。基于单晶LiNbO₃忆阻器组成的规模为9×3的忆阻器阵列实现了对热流传感器信号的二值神经网络边缘计算,进而实现了对3×3规模的热流传感阵列上三种不同字母形状的激光轨迹的识别。展现了基于忆阻器的二值神经网络在传感器端的神经网络边缘计算方面的应用潜力,为未来构建感-存-算一体系统奠定了基础。更多还原

【Abstract】 With the rapid development of intelligent Io T,artificial intelligence,and big data analytics,global data generation has exhibited exponential growth.In traditional data processing architectures,the von Neumann structure-based centralized processing mode requires sensing terminals to upload all collected data to data centers for processing,inevitably encountering"power wall"and"memory wall"challenges.Under this background,edge computing has attracted widespread attention,where memristors with non-volatile storage characteristics are regarded as key devices for breaking through edge computing bottlenecks due to their unique parallel computing capability,low-power characteristics,and high integration advantages.Compared to traditional sensors,thermal flow sensors demonstrate significant advantages in high-speed response and the direct generation of voltage pulses under laser excitation.However,in the implementation of memristor-based neural network edge computing for thermal flow sensing signal processing,challenges arise from the inherent non-ideal characteristics of memristive devices,mismatched output voltage pulse amplitudes,and peripheral circuit design complexities induced by intricate network topologies.Consequently,research on memristor-based thermal flow sensing signal processing remains largely unexplored.To address these issues,this study constructs a binary neural network to systematically characterize the non-ideal electrical properties of LiNbO₃ single-crystal memristors（high-resistance state resistance fluctuation and low-resistance state conductance continuous decay）.Through conductance-weight mapping models,we quantitatively analyze the impact of memristor characteristics on binary neural network recognition accuracy,ultimately verifying the strong robustness of binary neural networks against memristor non-idealities.Based on these findings,we establish a memristor-based binary verification platform for thermal flow sensors at the edge.By designing modular peripheral circuits to couple memristors with thermal flow sensors,functional validation is achieved through real-time classification of laser trajectories.Experimental verification using standard laser trajectories of letters"z","v",and"n"demonstrated the successful implementation of memristor-based binary neural network edge computing for thermal flow signal classification,achieving laser trajectory pattern recognition.The main research work is as follows:（1）Investigation of memristor non-ideal characteristics’impact on binary neural network recognition accuracy.Conducted research on binary neural network memristive characteristics using LiNbO₃ thin-film memristors.Established a 9×3 binary neural network trained with 3×3 pixel datasets of"z","v","n"characters.Focused on memristor data retention characteristics through Python-based modeling.Mapped non-ideal characteristics（LRS decay/HRS fluctuation）to neural network weight matrices.Tested recognition accuracy using standard and deformed characters to analyze non-ideality effects.（2）Research on binary neural network edge computing for thermal flow sensing signals using LiNbO₃ memristors.Investigated fundamental principles and design methods for thermal flow sensor-based BNN edge computing systems.Converted laser trajectories into sensing signals based on thermoelectric potential generation principles.Correlated memristor HRS/LRS states with binary weights（0/1）in neural networks.Explored coupling methods between thermal flow sensors and memristors.A 9×3memristor array composed of single-crystalline LiNbO₃ memristors was implemented to perform binary neural network（BNN）edge computing on thermal flow sensor signals,thereby achieving recognition of laser-induced thermal patterns corresponding to three distinct alphabetical characters on a 3×3 thermal sensor array.This work demonstrates the application potential of memristor-based BNNs in sensor-side neural network edge computing,laying the groundwork for future development of sensing-memory-computing integrated systems.更多还原