摘要
该文提出了一种基于麻雀搜索算法结合深度前馈神经网络(SSA-DFN)的近红外光谱模型转移方法。使用深度前馈神经网络拟合不同仪器采集到的光谱之间的非线性函数映射,并将麻雀搜索算法用于网络各层连接权值和阈值的初始化,通过种群中个体位置的迭代更新,求得连接权值和阈值的最优初始值;通过多次调整深度前馈神经网络模型的超参数,使网络拟合效果趋于最优,最终确定转移函数。为验证方法的有效性,分别从烟叶近红外光谱谱图、主成分投影和预测结果的角度,将SSA-DFN方法与分段直接校正算法(PDS)、典型相关性分析算法(CCA)转移前后的效果进行了对比。结果表明SSA-DFN方法转移后的从机光谱与原主机光谱重合度最高,转移后主、从机总糖、烟碱含量的预测结果差异不显著,预测平均误差从8.32%、9.15%分别降至4.65%、4.82%,预测均方根误差(RMSEP)和决定系数(R^(2))等指标均优于PDS和CCA,取得了最佳的转移效果,可满足企业需求。结果表明该方法是一种有效的模型转移方法。
In order to enhance the adaptability of the near-infrared model,a near-infrared spectral model transfer method based on sparrow search algorithm combined with deep feedforward neural net⁃work(SSA-DFN)was proposed in this paper,aiming at the problem that the function transfer rela⁃tionship between master and slave is difficult to determine due to the nonlinear interference caused by the difference between different stations and environmental factors.The depth feedforward network was used to fit the nonlinear function mapping between spectra collected with different instruments,and the sparrow search algorithm was used to initialize the connection weights and thresholds of each layer of the network.The optimal initial values of the connection weights and thresholds were ob⁃tained through iterative updating of individual positions in the population.By adjusting the super pa⁃rameters of the depth feedforward neural network model many times,the network fitting effect tends to be optimal,and finally the transfer function is determined.To verify the effectiveness of this meth⁃od,SSA-DFN method was compared with PDS and CCA before and after the transfer from the per⁃spectives of near infrared spectrum,principal component projection and prediction results.The re⁃sults showed that SSA-DFN method has the highest coincidence degree between the slave spectrum after transfer and the original host spectrum,and there is no significant difference in the prediction results of total sugar and nicotine content between the master and slave after transfer.The average prediction error decreased from 8.32%and 9.15%to 4.65%and 4.82%,respectively.The indexes such as RMSEP and R^(2) were better than PDS and CCA,which achieved the best transfer effect and met the needs of enterprises,indicating that this method is an effective model transfer method.
作者
刘鑫鹏
秦玉华
张凤梅
蒋薇
尹志豇
LIU Xin-peng;QIN Yu-hua;ZHANG Feng-mei;JIANG Wei;YIN Zhi-jiang(College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China;Technical Research Center,China Tobacco Yunnan Industrial Co.,Ltd.,Kunming 650231,China)
出处
《分析测试学报》
CAS
CSCD
北大核心
2022年第11期1621-1628,共8页
Journal of Instrumental Analysis
基金
云南中烟工业有限责任公司项目(2021JC03)。
关键词
模型转移
麻雀搜索算法
深度前馈神经网络
近红外光谱
calibration transfer
sparrow search algorithm
deep feedforward neural network
near infrared spectrum
