基于近红外光谱数据的一维卷积神经网络模型研究  

Study on One-Dimensional Convolutional Neural Network Model Based on Near-Infrared Spectroscopy Data

作　　者：唐杰 罗彦波[2] 李翔宇[2] 陈云璨 王鹏 卢天 纪晓波[4] 庞永强[2] 朱立军 TANG Jie;LUO Yan-bo;LI Xiang-yu;CHEN Yun-can;WANG Peng;LU Tian;JI Xiao-bo;PANG Yong-qiang;ZHU Li-jun(Chongqing Key Laboratory of Scientific Utilization of Tobacco Resources,Chongqing 400060,China;China National Tobacco Quality Supervision&Test Center,Zhengzhou 450001,China;Shanghai Shuzhiwei Information Technology Co.,Ltd.,Shanghai 200444,China;Department of Chemistry of Shanghai University,Shanghai 200444,China)

机构地区：[1]烟叶资源科学利用重庆市重点实验室,重庆400060 [2]国家烟草质量监督检验中心,河南郑州450001 [3]上海数之微信息科技有限公司,上海200444 [4]上海大学化学系,上海200444

出　　处：《光谱学与光谱分析》2024年第3期731-736,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(52102140);重庆中烟工业有限责任公司科技项目(HX202009)资助。

摘　　要：近红外光谱技术已被广泛应用于各种检测行业,但传统方法难以汇集光谱关键信息,导致模型预测误差较大。为减少误差,基于452个茄科植物,以化学成分为目标,探索了一维卷积神经网络(1DCNN)在近红外数据上的回归模型研究。经参数优化,总结了一套兼顾精度与训练效率的1DCNN模型参数,为后续模型研究提供参考。模型测试集的均方根误差为0.02~0.49,平均相对误差为0.8%~1.7%,远小于历史文献。相比传统方法,1DCNN可充分利用全部近红外谱图数据,且建模简单,模型预测能力强。该工作能为近红外光谱相关研究提供新的数据处理思路,也能促进该技术的应用与发展。Near-infrared spectroscopy technology has been widely applied for detection in various industries.However,traditional methods struggle to gather key information from the spectral data,leading to significant model prediction errors.This study explores the regression modeling of one-dimensional convolutional neural networks(1DCNN)on near-infrared data,focusing on the chemical composition of 452 plants from the Solanaceae family.Through parameter optimization,the study suggests that the optimal settings for the model include 64 channels in the intermediate convolutional layer,a maximum pooling layer of 1,6 convolutional layers,and 5 channels in the final convolutional layer.These findings can serve as a reference for future model research.The root mean square error of the model s test set ranges from 0.02 to 0.49,with an average relative error of 0.8%~1.7%,significantly lower than previous literature.Compared to traditional methods,1DCNN can fully utilize all of the near-infrared spectral data while maintaining a simple model structure and strong predictive capabilities.This work provides new insights for data processing in near-infrared spectroscopy research and promotes the application and development of this technology.

关 键 词：一维卷积神经网络 近红外光谱 深度学习 

分 类 号：O657.3[理学—分析化学]