基于神经网络的拉曼光谱波长选择方法  

Raman Spectrum Wavelength Selection Method Based on Neural Network

作　　者：沈东旭 洪明坚[1] 董家林 SHEN Dong-xu;HONG Ming-jian;DONG Jia-lin(School of Big Data&Software Engineering,Chongqing University,Chongqing 401331,China)

机构地区：[1]重庆大学大数据与软件学院,重庆401331

出　　处：《光谱学与光谱分析》2020年第11期3457-3462,共6页Spectroscopy and Spectral Analysis

基　　金：国家重点研发计划项目(2018YFF01011204)资助。

摘　　要：血液鉴别对于检验检疫、刑侦以及动物保护领域具有非常重要的意义,传统的的血液鉴别方法在鉴别的过程中存在分析周期长、对血液样本造成损害等缺点。而拉曼光谱可以通过分析与入射光频率不同的散射光谱得到分子振动、转动方面的信息,进而得到物质的组成成分,并且具有零污染非接触的特点,为血液的无损鉴别提供了可能,但是在拉曼光谱中,各个波长点之间存在严重的多重共线性,直接使用全光谱进行建模会增加模型的复杂性和降低模型的稳定性。针对拉曼光谱的特点,提出了一种基于神经网络的波长选择方法。该方法利用神经网络学习到各个波长点对校正模型的贡献权重,并将权重的均值作为阈值,去除权重低于阈值的波长点,以达到波长选择的目的。为了更容易确定筛选的阈值,在权重学习的过程中加入了稀疏约束,极大的减少了用于筛选的波长点。利用动物与人血清的拉曼光谱数据集对所提方法进行了验证,实验结果表明,利用该方法得到的光谱建立的校正模型,相比于全光谱数据在分类准确率和AUC值都有一定的提升,人工神经网络(NN)的准确率达到了94.495%, AUC值达到了0.9850,偏最小二乘(PLS-DA)的准确率达到了92.661%, AUC值达到了0.9760。与传统的波长选择方法UVE相比,该方法选择的波长点更少,仅选择了42个波长点用于建模,而且得到的校正模型的分类准确率和AUC值更高,证明该波长选择方法能有效的筛选出对建模有贡献的波长点,提高了模型的分类准确率和稳定性,为血液的无损鉴别提供了可能,具有一定的实用价值。Blood identification is crucial for the field of inspection and quarantine,criminal investigation and animal protection.Traditional blood identification methods have shortcomings such as long analysis period and damaging to blood samples during the identification process.Raman spectroscopy can obtain molecular vibration and rotation information by analyzing the scattering spectrum different from the incident light frequency,and obtain the composition of the material.Moreover,it has the characteristics of zero pollution and non-contact,which provides the possibility of non-destructive identification of blood.However,there is serious multicollinearity between each wavelength point in Raman spectrum,and directing the use of full-spectrum for modeling will increase the complexity of the model and reduce the stability of the model.According to the characteristics of Raman spectroscopy,this paper proposes a wavelength selection based on neural network.The method uses the neural network to learn the contribution weight of each wavelength point to the correction model,and uses the mean value of the weight as the threshold value to remove the wavelength point whose weight is lower than the threshold value,so as to achieve the purpose of wavelength selection.In order to make it easier to determine the threshold of the screening,sparse constraints are added to the weight learning process,which greatly reduces the wavelength points used for screening.The proposed method was validated by Raman spectroscopy datasets of animal and human serum.The experimental results show that the model established by the wavelength selection using this method has a certain improvement in classification accuracy compared and AUC value with the full spectrum,the accuracy of artificial neural network(NN)reached 94.495%and AUC value reached 0.9850.The accuracy of PLS-DA reached 92.661%,and AUC value reached 0.9760.Compared with the traditional wavelength selection method UVE,the method selects fewer wavelength points,and only 42 wavelength points

关 键 词：光谱分析 拉曼光谱 波长选择 神经网络 校正模型 

分 类 号：O657.37[理学—分析化学]