[Online Analysis of Dynamic Trend Regression and Endpoint Determination for Chinese Traditional Medicine Extraction Process Based on Ultraviolet spectroscopy].

With the transition of Chinese traditional medicine manufacture industry, modernization has become the inexorable tendency in its future development. However, during the current Chinese traditional medicine producing process, the lack of online monitoring leads to the lagging of quality detection, as well as quality differences between products. In this paper, aiming at realizing online monitoring and end point automatic determination for Chinese traditional medicine (CTM) extraction unit, which is one of the most important units in CTM producing, ultraviolet (UV) spectroscopy is applied to build UV absorbance dynamic model based on the Lambert-Beer law and the CTM extraction kinetic model which presents a new method of UV absorbance dynamic analysis and endpoint determination, including curve regression, robustness analysis and endpoint calculation. In the experiment for online monitoring homalomena occulta extraction, the online UV spectral collection system for CTM extraction, developed by our laboratory, was applied for spectral collection; meanwhile, solid component contents in the offline samples were measured as reference. During the analysis, first of all, we pretreated the spectrum collected in the current period with interpolation and smoothing, and calculated the mean value within a UV region of 230.2～400 nm to form an absorbance mean value sequence with data obtained in the early measurements; then we verified the linear correlation between the sequence of absorbance and the concentrations of effective component in the solution, the linear correlation coefficient equals 0.982 8, showing a high linearity between UV spectra and solid component contents; finally, we regressed the absorbance mean value sequence with the dynamic model, analyzing its robustness and the extraction endpoint. Experimental results demonstrate that, the robustness analysis could recognize the bad points of measurement during the regression process, and improve the consistency between the regression and the original curves, raising its squared correlation coefficient to more than 0.99; meanwhile, with endpoint determination, we shortened the homalomena occulta extraction process from the original manually set 180 to 122 min effectively. The experiment above proves that this method with UV spectroscopy realizes online monitoring and automatic endpoint determination for the CTM extraction process, and is of significant importance for stabilizing production as well as improving economic benefit.

Raman Spectral Analysis of Low-content Benzene Concentration in Gasoline with Partial Least Squares Based on Interference Peak Subtraction.

Raman spectroscopy is adopted to detect the low-content benzene concentrations in gasoline products. Due to the peak overlap of benzene and other species in the gasoline spectrum, the associated statistical regression methods cannot make stable predictions unless there are enough training samples. To extend their extrapolation to small-size training sets, we propose the method of partial least squares based on a spectral pretreatment of interference peak subtraction (IPS-PLS). During the analysis, after spectral interpolation and baseline removal, we extract the benzene peak by interference peak subtraction (IPS), and then partial least squares (PLS) is applied to make a prediction. The experimental results demonstrate that, IPS can extract benzene information effectively, and help to decrease principal components needed by PLS, thus IPS-PLS is superior to direct PLS with small-size training sets, and depends less on the training sample distribution. Meanwhile, IPS-PLS can reach the standard of ASTM 3606-10 with the least of 9 training samples, while keeping its max predictive error less than 0.1254% (v/v), which shows promising prospects in gasoline quality test.

[Component analysis of the circulating fluid in an adsorption tower in a P-xylene unit based on raman spectral decomposition].

In order to achieve fast and accurate online analysis of the circulating fluid in an adsorption tower in a p-xylene unit, the Raman spectral analysis method is adopted. However, the Raman spectra of the pure components included in the circulating fluid overlap together, and the concentration of each component varies obviously, the present Raman analysis technology needs a large amount of training samples. Therefore, this paper applies Raman spectral decomposition method in component analysis of the circulating fluid. First of all, the Raman spectra of the pure components and the spectra of a few training samples must be measured, and baseline subtraction and mean normalization are applied to obtain pretreated Raman spectra. Then the characteristic wave number range, 680-880 cm(-1), is chosen, and the Raman spectral decomposition method is adopted, to get decomposition coefficients of each component for each training sample. Next, the mathematical model between coefficients and concentrations of each component are built based on all training samples. For a testing sample, the above spectral pretreatment and the spectral decomposition for the same wave number range is adopted, then the decomposition coefficients of each component can be obtained. Based on the built mathematical model, the concentrations of all components can be predicted. Experimental results show that the standard prediction errors for the concentration of toluene, ethylbenzene, p-xylene, m-xylene, o-xylene and p-diethylbenzene are 0.301%, 0.088%, 0.563%, 0.384%, 0.366% and 0.536% respectively. The above method provides a methodological basis for the online analysis of the circulating fluid in adsorption towers.

[Fast discrimination of edible vegetable oil based on Raman spectroscopy].

A novel method to fast discriminate edible vegetable oils by Raman spectroscopy is presented. The training set is composed of different edible vegetable oils with known classes. Based on their original Raman spectra, baseline correction and normalization were applied to obtain standard spectra. Two characteristic peaks describing the unsaturated degree of vegetable oil were selected as feature vectors; then the centers of all classes were calculated. For an edible vegetable oil with unknown class, the same pretreatment and feature extraction methods were used. The Euclidian distances between the feature vector of the unknown sample and the center of each class were calculated, and the class of the unknown sample was finally determined by the minimum distance. For 43 edible vegetable oil samples from seven different classes, experimental results show that the clustering effect of each class was more obvious and the class distance was much larger with the new feature extraction method compared with PCA. The above classification model can be applied to discriminate unknown edible vegetable oils rapidly and accurately.

[Fast spectral modeling based on Voigt peaks].

Indirect hard modeling (IHM) is a recently introduced method for quantitative spectral analysis, which was applied to the analysis of nonlinear relation between mixture spectrum and component concentration. In addition, IHM is an effectual technology for the analysis of components of mixture with molecular interactions and strongly overlapping bands. Before the establishment of regression model, IHM needs to model the measured spectrum as a sum of Voigt peaks. The precision of the spectral model has immediate impact on the accuracy of the regression model. A spectrum often includes dozens or even hundreds of Voigt peaks, which mean that spectral modeling is a optimization problem with high dimensionality in fact. So, large operation overhead is needed and the solution would not be numerically unique due to the ill-condition of the optimization problem. An improved spectral modeling method is presented in the present paper, which reduces the dimensionality of optimization problem by determining the overlapped peaks in spectrum. Experimental results show that the spectral modeling based on the new method is more accurate and needs much shorter running time than conventional method.

[The application of piecewise direct standardization with SNV in calibration transfer of Raman spectra].

To implement calibration transfer between Raman spectrometers, an improved piecewise direct standardization (PDS) is proposed in the present paper. Standard normal variate (SNV) is firstly introduced to reduce the influence of spectral background and intensity corresponding to the master spectrometer and the slave spectrometer; then PDS algorithm is used to eliminate the differences between Raman spectra for a specific sample. Moreover, a new quantitative criterion, called transfer error rate, is proposed to evaluate the performance of calibration model transfer. This improved PDS is applied to Raman spectral analysis of gasoline. The result shows that the proposed algorithm not only needs a small quantity of transfer samples, but also obtains high transfer accuracy and strong model robustness.

[A fast classification method for petroleum products based on the Raman spectroscopy].

A fast and effective method for classification of petroleum products based on Raman spectroscopy is proposed. A knowledge base composed by Raman spectra of training samples, intra-class feature spectra and intra-class thresholds of all classes was firstly established. Then, correlation coefficients between the test sample and the intra-class feature spectra were calculated. If the maximal correlation coefficient of the test sample is larger than or equal to the corresponding intra-class threshold, the test sample is determined to belong to the corresponding class. For 96 petroleum product samples belonging to 7 classes and 4 unknown samples, the experimental results show that this method can accurately classify known test samples and can also find the unknown test samples. This method costs little calculation time and human interference. Moreover, it can be easily implemented in the practical application.

[Application of Raman spectra feature extraction in chemical fiber component qualitative identification].

According to the characteristics of the textile fibers Raman spectra, a qualitative identification method based on Raman feature extraction is proposed. This fast method consists of spectrum measurement and spectral data processing algorithm, including spectrum preprocessing, feature extraction and matching recognition. It can be used to identify the components of fibers or fabrics, especially chemical fibers, which is an inspective difficulty in daily analytic work for its remarkable Raman feature. The authors performed an experiment to analyze 4 typical and widely used kinds of fibers as algorithm verification. They are terylene fiber, acrylic fiber, nylon fiber and rayon fiber. To identify the components of one test sample, first the authors set up feature tables of these 4 standard samples, which describe the features of their preprocessed spectra containing both position information and intensity information, then extract features of the test sample. The authors match these features with the tables and calculate the matching confidence coefficients of the results, which can be used to filter the unexpected matching results caused by accident and attain the final qualitative identification result. The experimental results confirm that this method is effective, efficient and expansible, which means it can be used to identify more actual fiber types by adding more standard spectra to the feature table database. In addition, it is a pure optical method, which needs only a small quantity of sample without any pretreatment. The whole identification process is damage-free, pollution-free and suitable for various kinds of fabrics. Compared to all existing methods, this Raman spectrum identification method can solve the limitation of efficiency, pollution, universality, and fill a gap in fabric inspection field.

[Hydrocarbon group analysis of gasoline by Raman spectroscopy with outlier detection].

Hydrocarbon group of gasoline not only determines the quality of gasoline, but also directly relates to the impact of exhaust air on the environment. The present paper successfully applied Raman spectroscopy to the quantitative analysis of hydrocarbon group in gasoline. Contaminated samples were removed from calibration set by outlier detection, which effectively improved the partial least squares (PLS) prediction accuracy. The standard prediction errors of aromatics, olefin and oxygen content are 0.23, 0.52 and 0.143 respectively, and the corresponding multiple correlation coefficients of prediction (R2) are 0.987, 0.927 and 0.971. Experimental results show that the effectiveness of Raman spectroscopy for hydrocarbon group analysis in gasoline, and that the prediction accuracy is much better than near infrared spectroscopy or multidimensional gas chromatography. Quantitative calibrations based on Raman spectroscopy can also be used in the on-line analysis of gasoline production processes.

[Fast recognition of gasoline brands based on the Raman spectroscopy].

A novel method for fast recognition of gasoline brands based on the Raman spectroscopy is presented. A classification model on the basis of product gasoline samples with known brands was established. The detailed modeling process includes measurement and pretreatment of Raman spectra of these samples, principal component analysis (PCA) to obtain loading vectors and score vectors of all known samples, and calculating each average score vector for all of the samples with the same brand. For a gasoline sample with unknown brand, first measure and preprocess its Raman spectrum with the same pretreatment algorithm, then calculate its score vector on the above loading vectors and its distances to the average score vectors for different brands, and finally determine the brand of the unknown sample by the minimum distance. For 45 product gasoline samples from different refinery, experimental results show that there are significant distances between different brands in the principal component space, and the above classification model can decide the brand of unknown gasoline samples rapidly and accurately.

[Fast analysis of benzene concentration in gasoline based on low-resolution dispersive Raman spectroscopy].

In order to fast analyze the benzene concentration in gasoline, a new measure method based on low-resolution dispersive Raman spectroscopy is proposed. There exist strong measurement noise and fluorescence background in dispersive Raman spectra, so the present paper applies the Savitzky-Golay smoothing filter to remove the measurement noise and uses iterative polynomial curve-fitting to reduce the fluorescence background. Based on ridge regression, principal component regression and partial least squares algorithm, three calibration models of the benzene concentration in gasoline are built and validated by a set of gasoline samples from a refinery. Experimental results show that their repeatability and reproducibility can satisfy the accuracy requirement specified by the standards SH/T0713-2002, regardless of applying what kind of calibration models. In addition to its low cost, small size, convenience to use and so on, the fast measure method based on low-resolution dispersive Raman spectroscopy can be widely applied to the routine analysis.

[Spectral wavelength selection based on PLS projection analysis].

In order to enhance the prediction accuracy of spectral analysis models and reduce their input number, this paper presents a simple and rapid wavelength selection method based on PLS projection correlation coefficients. These correlation coefficients are decided by both the changes in spectra data and the PLS regression coefficients between spectra matrix and concentration vector. Compared with the traditional wavelength selection method based on correlation analysis, the novel proposed method obviously improves the robustness of spectral analysis models and reduces their input number sharply. Applying the proposed method to 208 gasoline samples, the experimental results show that the number of calibration model input decreases to 30% of the original wavelength number, and the root mean square error of cross validation is reduced from 0.44 to 0.34. This method can be widely used in wavelength selection and data compression in spectral quantitative analysis.

[A simple and convenient standardization algorithm of NIR spectra].

Due to the limitation of current algorithms for NIR spectral analysis model transfer, a simple and convenient algorithm to standardize the spectra was proposed, and a new performance index called spectra standard error (SSE) was also constructed to evaluate the validity of model transfer algorithms. SSE expresses the ratio of J2 to J1, where J2 describes the distances between the spectra of the same sample using different instruments, and J2 describes the average distance between the spectra of different samples using the original instrument for their central spectrum. In the present paper we first used Savitzky-Golay smoothing to realize baseline correction for different spectra, and then applied standard normal variate method to standardize spectra and polynomial filtering to avoid noise. Besides, we optimized the wavelength range and the window width in Savitzky-Golay smoothing in order to minimize the SSE. After these steps, the standardized spectra can be applied to spectral analysis modeling. By the new algorithm, neither collecting a great number of samples nor need measuring all spectra of training samples using different instruments are needed. For a set of gasoline samples, SSE can be reduced from 1.418 to 0.167 via the new standardization algorithm and satisfactory results were obtained.

[Near-infrared spectral quantitative analysis by combining classification with local PLS].

As a rapid analytical technology, near-infrared (NIR) spectroscopy has been developed fast in recent years. To improve the accuracy of near-infrared spectral quantitative analysis, the present paper first classifies a testing sample by a support vector machine classifier and selects some similar training samples of the same type to build the calibration model, than predicts the property of the testing sample. To avoid the negative influence of classification failure, a new hybrid algorithm (called H_PLS) was proposed. This algorithm consists of a local PLS model based on the same-type training samples (called C_PLS) and a local PLS model based on the total training samples (called D_PLS). H_PLS calculates the predictive value of the property for the testing sample by comparing the outputs of the two models. For a set of gasoline samples, experimental results show that the prediction accuracy of C_PLS is higher than that of D_PLS if there are no classification errors, otherwise the prediction accuracy of C_PLS will drop obviously. The novel proposed algorithm (H_PLS) combines the advantages of C_PLS and D_PLS. Using H_PLS, can increase from 0.9734 of D_PLS and 0.9656 of C_PLS to 0.9858 even though there are classification errors.