卷积桥接孪生自编码器的近红外光谱转移研究  

Study on Near Infrared Spectrum Transfer of Convolutional Bridged Twin Denoising Reduction Autoencoder

作　　者：杨泽会 吴箭 李瑞东 郝贤伟[3] 吕小芳 田雨农 张志成 吴灵通 李正莹 夏春艳 张恺 徐梦瑶 毕一鸣[3] 夏自麟 YANG Ze-hui;WU Jian;LI Rui-dong;HAO Xian-wei;LU Xiao-fang;TIAN Yu-nong;ZHANG Zhi-cheng;WU Ling-tong;LI Zheng-ying;XIA Chun-yan;ZHANG Kai;XU Meng-yao;BI Yi-ming;XIA Zi-lin(Xuanwei Redrying Factory of Yunnan Tobacco Redrying Co.,Ltd.,Xuanwei 655400,China;Technical Center of Yunnan Tobacco Redrying Co.,Ltd.,Kunming 650000,China;Technical Center of Zhejiang Zhongyan Industry Co.,Ltd.,Hangzhou 310024,China;Material Department of Zhejiang Zhongyan Industry Co.,Ltd.,Hangzhou 310002,China;Yunnan Mingfan Technology Co.,Ltd.,Kunming 650051,China)

机构地区：[1]云南烟叶复烤有限责任公司宣威复烤厂,云南宣威655400 [2]云南烟叶复烤有限责任公司技术中心,云南昆明650000 [3]浙江中烟工业有限责任公司技术中心,浙江杭州310024 [4]浙江中烟工业有限责任公司物资部,浙江杭州310002 [5]云南铭帆科技有限公司,云南昆明650051

出　　处：《分析测试学报》2025年第3期471-478,共8页Journal of Instrumental Analysis

基　　金：云南烟叶复烤有限责任公司科技计划项目(2022FK06);中国烟草总公司浙江中烟工业有限责任公司科技计划项目(ZJZY2023A012)。

摘　　要：近红外光谱仪器间的差异使得不同仪器共用预测模型变得困难,限制了技术的推广应用。为减少光谱偏移后重新建立预测模型的难度,该文提出了一种基于卷积桥接孪生降噪自编码器(CBSDAE)的近红外光谱模型转移方法。该方法利用卷积降噪自编码器(CDAE)的编码器提取光谱的隐藏特征,并通过卷积神经网络拟合从机与源机光谱隐藏特征的转移映射函数,最后通过卷积降噪自编码器的解码器重构转移后的光谱。为验证其有效性,该文对烟叶近红外光谱图及化学成分预测结果进行评估。结果显示,CBSDAE方法转移后的从机光谱与源机光谱高度重合。相比于直接标准化(DS)、分段直接标准化(PDS)、光谱差值校正算法(SSC)、Shenk’s算法、卷积神经网络(CNN)、深度自编码器算法,使用该方法进行光谱转移后,预测烟碱的平均相对误差分别下降了6.42%、5.84%、5.32%、5.24%、4.35%和4.85%,预测的均方根误差(RMSEP)和相关系数也均优于上述方法。结果表明该方法是一种有效的模型转移方法。The differences between near infrared(NIR)spectrometers make it challenging to share prediction models across different instruments,limiting the widespread application of this technology.To reduce the difficulty of rebuilding prediction models after spectral shift,this paper proposes a near infrared spectral model transfer method based on a convolutional bridged twin denoising autoencoder(CBSDAE).This method utilizes the encoder of the convolutional denoising autoencoder(CDAE)to extract the hidden features of the spectra and fits a transfer mapping function between the hidden spectral features of the target and source instruments using a convolutional neural network(CNN).Finally,the transferred spectra are reconstructed through the decoder of the CDAE.To validate its effectiveness,evaluations were conducted from two perspectives:NIR spectra of tobacco leaves and chemical component prediction results.The findings show that the spectra from the target instrument closely overlap with those from the source instrument after transfer using the CBSDAE method.Compared with direct standardization(DS),piecewise direct standardization(PDS),spectral subtraction correction(SSC),Shenk’s algorithm,CNN and deep autoencoder,the average relative error in nicotine prediction decreased by 6.42%,5.84%,5.32%,5.24%,4.35%and 4.85%,respectively,after applying the CBSDAE method for spectral transfer.Additionally,the root mean square error of prediction(RMSEP)and correlation coefficient were superior to those of the aforementioned methods.These results indicate that the proposed method is an effective approach for model transfer.

关 键 词：模型转移 编码器 孪生 卷积桥接 近红外光谱 

分 类 号：O657.3[理学—分析化学] F768.29[理学—化学]