复杂体系高灵敏拉曼定量分析中的多尺度建模方法

Rapid non-destructive detection of trace components in complex systems represents one of the challenges in analytical chemistry. Liquid core waveguide Raman spectroscopy, a novel high-sensitive Raman spectrometry technology, has the advantages of rapidness and non-destruction as well, thus possessing the potential of rapid trace analysis. However, the model of high-sensitive Raman spectra collected from complex system is determined not only by the concentrations of analytes, but also by the matrix interference, nonlinear effects and sample fluctuation presented in complex systems as well, making the conventional algorithms difficult to solve these interference simultaneously.In this regard, this project begins with the model driven multiscale modeling to solve the spectral and non-linear interference using the combined strategies of multiscale decomposition in time-frequency domains and data fusion. Then a novel method based on data driven is proposed to solve the sample fluctuation using local approximation. Finally, a novel strategy of dual driven multiscale modeling (DDMM) is established by integrating the advantages of model-driven global optimization and data-driven local approximation, and thus suppressing the interference mentioned above. DDMM fully utilizes the multiscale characteristics of time-frequency domains in spectra to prevent from information leakage through the integration of signal preprocessing and multivariate calibration, extracting the information of multiple trace components from high-sensitive Raman spectra adaptively. As a result, DDMM can greatly improve the reliability and prediction precision of Raman quantitative analysis, which is expected to provide an efficiently novel method for complex signal resolution, thus promoting the applications of chemometrics.

对复杂体系微量成分进行快速无损检测，一直是分析化学所面临的挑战之一。液芯波导拉曼光谱作为一种新型高灵敏光谱技术，以其快速、无损等优点，具备微量速测的潜力。但对复杂体系的高灵敏拉曼光谱进行建模时，其模型不仅取决于待测组分浓度，还受光谱干扰、非线性效应、样品波动性等因素的影响，常规算法难以同时解决相关干扰。针对以上问题，本项目从模型驱动的多尺度模型入手，借助时/频多尺度分解和数据融合策略，解决光谱和非线性干扰；提出数据驱动新思路，以局部逼近的方式解决波动干扰。最后联合模型驱动全局优化和数据驱动局部逼近的优点，创建双驱动多尺度建模方法，同时克服相关干扰。双驱动多尺度算法充分利用光谱的时/频多尺度特性，通过数据预处理与多元校正的一体化运算避免了信息丢失，自适应地提取高灵敏拉曼光谱中的微量多组分信息，有效提高其定量分析的可靠性及预测精度，从而为复杂信号解析提供一种新手段，并促进化学计量学的推广应用。

项目以复杂体系中微量多组分检测为典型例子，发展模型驱动和数据驱动的多尺度建模新方法，由此创建双驱动多尺度模型以实现复杂信号的准确解析，进而高效满足高灵敏拉曼光谱的定性定量分析需求。项目首先从模型驱动的多尺度建模方法入手，创新性引入双树双密度小波、自适应小波等第二代小波变换新算法，着力改善多尺度模型的时/频多尺度分解和数据融合性能，并在此基础上发展了一种通用非线性变量筛选新方法，有效解决了光谱干扰和非线性响应问题。然后提出数据驱动的新思路，以局部逼近的方式解决光谱波动因素，并应用于奶粉、食用油等食品掺假的快速筛选中，可自适应地提取拉曼光谱信号中的掺假物资信息并加以甄别，其识别准确率优于98%。最后联合模型驱动全局优化和数据驱动局部逼近的优点，发展基于二维相关光谱的双驱动多尺度模型，显著提升了拉曼光谱技术对复杂体系中重叠峰的分辨能力，并成功应用于橄榄油掺杂识别中，有望在复杂体系光谱分析中得到广泛的应用。