【作者】 姜承志；

【导师】 梁静秋； 孙强；

【作者基本信息】 中国科学院研究生院(长春光学精密机械与物理研究所) ， 光学， 2014， 博士

【摘要】 拉曼光谱分析技术由于具有无损、信息丰富、无需样品制备等优点，在食品、材料、环境监测等众多领域得到了越来越广泛的应用。手持拉曼光谱仪由于具有操作简单、小巧轻便等优点被广泛应用于工业生产中的物料鉴定。目前，国外各大光谱仪生产商均已经推出了各种型号的手持拉曼光谱仪，国内市场已被其垄断，因此研制我国拥有独立自主知识产权的手持拉曼光谱仪具有重要意义。本文针对手持拉曼光谱仪数据处理与定性分析技术展开了相关研究。由于手持拉曼光谱仪主要应用于工业生产中的定性判别问题，生产线上操作员往往不具备专业的化学分析知识，因此减少拉曼光谱分析过程中的人工干预，实现拉曼光谱数据处理与定性分析的自动化是手持拉曼光谱定性分析技术的关键。本文系统地研究了手持拉曼光谱的数据处理与定性分析算法流程，主要研究工作如下：(1)研究了拉曼光谱质量评价方法。实现了小尺度小波变换法、空域相关小波变化法、Donoho鲁棒估计法及改进的二阶差分法等噪声标准差估计方法，对比了其估计噪声标准差的精度，结果表明改进的二阶差分法的估计精度最高；提出了一种新的信噪比计算方法，新的信噪比计算方法与传统的信噪比定义相比能更好的表征信号质量。(2)研究了拉曼光谱数据处理技术。实现了各种常见的光谱预处理方法，重点研究了拉曼光谱的去尖峰、去基线、去噪声方法。提出了一种无需设置任何参数的、可实现完全自动化的去尖峰方法——改进的循环消去法；实现了一种基于三点零阶Savitzky-Golay滤波器的自动化降噪方法，对比了其与传统的滑动窗口平均法、滑动窗口中位值法、Savitzky-Golay滤波器法、小波阈值滤波法的降噪效果，数值实验表明该方法具有最优的去噪效果，同时引起的谱峰退化程度也最小，能够最大量的保留谱峰信息；提出了一种新的基线估计方法——改进的小窗口滑动平均法，该方法无需设置任何参数，可实现基线的自动估计，估计精度良好。(3)研究了拉曼光谱谱峰识别技术。实现了连续小波变换法识别拉曼谱峰，提出了两种新的拉曼谱峰识别方法——双尺度相关法及多尺度局部信噪比法，对比了其与连续小波变换法识别拉曼谱峰的能力。仿真实验表明，多尺度局部信噪比法具有最优的谱峰识别能力。对于处于检测限的单峰，仍有95.1%的识别准确率，谱峰信噪比大于等于6时谱峰识别准确率高达100%；对于重叠峰，谱峰信噪比大于等于7时达到100%识别。多尺度局部信噪比法具有最高的峰位估计准确度。(4)研究了拉曼光谱判别分析技术。实现了直接比较法和基于簇类的软独立模型法，对比了两者的性能。直接比较法的不足是，当参考谱库中没有待测样品的参考光谱时，直接比较法仍然会给出一个最佳的匹配结果。基于簇类的软独立模型法具有更加优越的性能，该方法具有较高的识别准确率，当未知样品不属于类库中的任何类时，该方法可识别出未知样品属于某一新的类型。利用基于簇类的软独立模型法还可以获得两类间的相似度、变量对样品判别的重要性、样品与某类的相关性等信息。更多还原

【Abstract】 Raman spectrum analysis technique is gaining a wider use in many fields suchas food, materials and environmental monitoring due to its advantages ofnon-destructive, rich information, no sample preparation etc.. Handheld Ramanspectrometer is widely used in the material identification of industrial productionbecause of its advantages of easy to operate, compact construction, lightweight etc.Currently, major foreign spectrometer manufacturers have already launched variousmodels of handheld Raman spectrometer. The domestic market has beenmonopolized by these products. Therefore the need of development of our ownhandheld Raman spectrometer with independent intellectual property rights isbecoming extremely urgent.In this paper, related research on data processing and qualitative analysis ofhandheld Raman spectrometer is carried out. Because handheld Raman spectrometeris mainly used in qualitative discrimination issues in industrial production, andproduction line operators often do not have the professional knowledge of chemicalanalysis, reducing the need of manual intervention and achievement of automation ofdata processing and spectrum analysis process become a critical point.This paper systematically studied the algorithm of data processing andqualitative analysis of handheld Raman spectroscopy. The major research work is as follows:(1) Studied the quality evaluation methods of Raman spectrum. Noise standarddeviation methods such as small scale wavelet transform method, spatial correlationwavelet transform method, Donoho robust estimation method and improvedsecond-order difference method were realized. Estimation accuracy of noise standarddeviation of these methods were compared, results showed that improvedsecond-order difference method is the most accurate estimation method. Proposed anew method for the calculation of SNR, the new SNR can better characterize thequality of Raman spectrum in comparison with the conventional SNR.(2) Studied the data processing methods of Raman spectrum. The commonspectrum preprocessing methods were realized. The research were focused on thedespeike, debaseline and denoise methods of Raman spectrum. Proposed a novelmethod for spectrum despike, improved iteratively stripping method, which is atotally automated despike method without any parameter need to be specified.Realized an automated noise reduction method based on three points zero-orderSavitzky-Golay filter. It’s denoise performance was compared with the traditionaldenoise methods such as sliding window average, sliding window median method,Savitzky-Golay filter and wavelet threshold filter. Numerical experiments showedthat the method had the best denoising performance. Meanwhile the degradation ofpeak caused by denoising was minimal. It could retain most of the informationcontained in Raman peaks. Proposed a novel baseline estimation method, improvedsmall window moving average method, which is a totally automatic debaselinemethod without any parameters to be set. The novel method’s baseline estimationaccuracy is pretty good.(3) Studied the peak recognition methods of Raman spectrum. Realized the peakrecognition method based on continuous wavelet transform. Proposed two novel peakrecognition methods: bi-scale correlation method and multi-scale local SNR method.Their peak recognition performance of these methods was compared. Simulationresults showed that the multi-scale local SNR method outperformed the continuous wavelet transform method and bi-scale correlation method. The simulationexperiment shows that: for singular peak on the detection limit the recognitionaccuracy of multi-scale local SNR method is95.1%, when the peak SNR is greaterthan or equal to6, the recognition accuracy of proposed algorithm is up to100%, forcongested peak, when the peak SNR is greater than or equal to7, the recognitionaccuracy is up to100%, multi-scale local SNR method had the highest peak positionestimation accuracy.(4) Studied the discriminant analysis methods of Raman spectrum. Directcomparison method and soft independent modeling of class analogy method wererealized. Their performance was compared. The disadvantage of direct comparisonmethod is that when the reference spectrum of unkown sample wasn’t contained inthe reference library, the direct comparison method would still give one best matchresult. Soft independent modeling of class analogy method outperformed the directcomparison method. The soft independent modeling of class analogy method hadmore superior identification performance. When the reference spectrum of unkownsample wasn’t contained in the reference library the soft independent modeling ofclass analogy method could tell us that the unkown sample belonged to a new class.Through soft independent modeling of class analogy method we could get otherinformation such as the similarity between the two classes, the relevance of avariable and the similarity between one sample and certain class.更多还原