摘要
木材的抗弯强度是木材重要的力学指标。光谱分析操作简单、方便、快速,已成为木材检测的重要手段。但是在应用中,面对检测环境的温湿度变化、仪器部件老化和附件更换等情况,采集到的光谱数据会发生一定程度的偏移。为了解决这一问题,以落叶松抗弯强度的近红外光谱预测模型为研究对象,针对不同类型光谱仪数据差异而导致主机模型泛化能力差的问题,提出了一种迁移学习与光谱转移校准结合的近红外光谱建模方法。加工200组落叶松板材试件样本,以NIRQuest512光谱仪为主机、One-chip微型集成光谱仪为从机,分别采集落叶松试材光谱数据,利用力学万能测试机检测试件力学真值;2类数据经过SNV、SG、光谱剪切预处理后,从机数据使用PDS转移校正完成从机到主机光谱线性变换;然后,利用SWCSS对2类光谱数据进行特征提取,优选出主机与从机相关的稳定性一致光谱波段;最后,采用100组试件的2类近红外光谱数据进行GFK-SVM建模,得到适用于主机、从机设备的通用模型。为了验证模型方法的有效性,应用100组数据进行测试并对比了DS-PLS,PDS-PLS,DS-SWCSS-GFK-SVM和PDS-SWCSS-GFK-SVM等建模方法;结果表明,PDS采用了滑窗技术,方法相较DS方法可以更好地完成光谱数据的线形映射,在一定程度上提高了建模精度,统一了两组光谱仪之间的光程与波长数;SWCSS特征提取方法能够根据2组光谱数据集之间的差异与共性优选波段,保证特征选择的有效性与稳定性,提升建模精度;GFK-SVM适合不同光谱数据的迁移,能够通过合理的核函数参数优选实现不同类型光谱数据的高维映射,在高维空间中构建不同数据集的通用模型,实现主机模型在从机光谱预测上的泛化,提升了数据的使用效率,测试集相关系数R_(p)达到0.875,均方根误差RMSEP为11.975。
The modulus of elasticity is an important mechanical index of wood.The advantages of spectral analysis technology include a simple,convenient and fast operation process,which has become an important tool for wood testing.However,in practical applications,we often face changes in temperature and humidity of near-spectrometer testing conditions or aging of instrument components and replacement of accessories when the collected spectral data will be shifted.In order to solve this problem,this paper proposes a near-infrared spectral modeling method combining migration learning and spectral transfer calibration to address the poor generalization of the master model due to the difference data from different types of spectrometers,taking the near-infrared spectral prediction model of larch bending strength as the research object.Firstly,200 sets of Larch test specimens were processed.Two kinds of spectrometers,the NIRQuest512 spectrometer as the master instrument and the One-chip as the slave,were used to collect the spectral data of Larch test specimens respectively.And the true values of the test specimens were detected by the mechanical universal testing machine.Secondly,the preprocessing of SNV,S-G and spectral shearing was employed,and then the method of PDS transfer correction was applied to complete the linear transformation from the slave instrument to the master.Thirdly,the SWCSS was used to extract the features of two kinds of spectral data,and the stable wave points were optimized.Finally,the GFK-SVM model was established by using two types of near-infrared spectral data of 100 sets of specimens.100 sets of data were applied to test and compare the modeling methods such as DS-PLS,PDS-PLS,DS-SWCSS-GFK-SVM,and PDS-SWCSS-GFK-SVM.The experimental results show that PDS,compared with DS,can better complete the linear mapping of spectral data due to the sliding window,which could unify the optical length and wave points between the two spectrometers,and improve the modeling accuracy to a certain extent;As a feature extraction method,SWSS can select wavebands according to the differences and similarities of the two groups of spectral data sets,which can ensure the effectiveness and stability of features,and improve the modeling accuracy;The GFK-SVM is suitable for the migration of different spectral data.It can realize high-dimensional mapping of different types of spectral data through reasonable kernel function parameters.A generalized model for different datasets is constructed to realize the generalization of the master model on the slave spectral prediction,which improves the data efficiency,and the test set correlation coefficient R;reaches 0.875,and the root mean square error RMSEP is 11.975.
作者
陈金浩
蒋大鹏
张怡卓
王克奇
CHEN Jin-hao;JIANG Da-peng;ZHANG Yi-zhuo;WANG Ke-qi(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2022年第5期1471-1477,共7页
Spectroscopy and Spectral Analysis
基金
国家林业局948项目(2015-4-52)
中央高校基本科研业务费专项资金项目(2572017CB34)资助。
作者简介
陈金浩,1978年生,东北林业大学机电工程学院博士研究生.e-mail:jinhaochen@vip.sina.com;通讯作者:张怡卓,e-mail:zdhwkq@163.com;通讯作者:王克奇,e-mail:nefuzyz@163.com。
