[Successive projections algorithm and its application to selecting the wheat near-infrared spectral variables].

Successive projections algorithm combined with partial least squares regression, termed as SPA-PLS approach, was implemented as a novel variable selection approach to multivariate calibration. The proposed approach was applied to near-infrared reflectance data for analyzing moisture in wheat. The number of variables selected from 701 spectral variables was reduced to 16 by SPA, and the root mean squared error of prediction set (RMSEP) of the corresponding partial least squares regression models was decreased to 0.205 5% as well. The result indicates that the SPA-PLS approach by performing SPA prior to calibration not only can improve the model accuracy, but also decreases the number of spectral variables, so its resulting model becomes more concise. Moreover, as compared with genetic algorithm for wavelength selection, SPA is a deterministic search technique whose results are reproducible and it is more robust with respect to the choice of the validation set.

[Classification of pharmaceutical tablet with canonical variates analysis method in spectra analysis].

A novel classifier was constructed in the present paper by combination of an improved canonical variates analysis (ICVA) with Fish linear discriminant analysis (LDA). The resulting discrimination model based on this proposed approach (ICVA-LDA) was divided into two parts: the inner part that estimated the robust weight vector of canonical variates by linear partial least square algorithm and the outer part that built the LDA discrimination model by making use of the extracted canonical variates. The method utilized partial least squares regression as an engine for solving an eigenvector problem involving singular covariance matrices and the canonical variates were more relevant for discriminative purposes. Thus, the weight vectors found in the modified CVA method not only possessed the same properties as weight vectors of the standard CVA method, but also forced the discriminative information into the first fewer of canonical variates. The improved discrimination model was more concise and efficient in dealing with the problem of the effect sensitivity and numerous predictor variables with serious multicollinearity in the spectra data. Furthermore, in ICVA-LDA the interpretation could be performed with respect to the original high-dimensional data space. Finally, application to a four-group problem with near-infrared transmittance spectroscopy data consisting of 310 samples and 404 variables of the proposed ICVA-LDA approach was presented with comparison to the LDA combined with principal component analysis (PCA-LDA) and standard CVA-LDA methods. All the three discrimination models were validated using fivefold segmented cross-validation. The result demonstrates that the limitations of LDA were overcome with PLS algorithm and then the classification performance of LDA was improved by ICVA. This proposed approach can also be widely used in other fields for classification and discrimination of small samples and collinear data.

[Quantitative analysis of electronic absorption spectroscopy by piecewise orthogonal signal correction and partial least square].

Electronic absorption spectroscopy (EAS), as an indirect analytical technique, has been used to carry out quantitative analysis of unknown samples by establishing a model with calibration samples. Partial least squares (PLS), as a powerful technique for process modeling and multivariate statistical process control, has been widely used to establish this model. On account of the noise signal in the spectra, the signal preprocessing was often a necessary step in building reliable and robust multivariate calibrations. The main goal of preprocessing was to remove variation in the data that was irrelevant to modeling. Orthogonal signal correction (OSC) and related methods emerged as filtering techniques for various spectra in modern chemometrics literature. However, it was confirmed in recent work that preprocessing with OSC did not lead to any significant improvements in calibration models subsequently developed by means of PLS regression, except for merely reducing the number of latent variables in the PLS model by the number of OSC components removed. Now in our study, taking into account the local effect sensitivity and numerous predictor variables with serious multicollinearity of the spectra data, a novel PLS algorithm that embedded the OSC into the regression framework of the PLS, termed as POSC-PLS method, was implemented. It firstly applied the OSC technique to a set of selected spectra at an optimized size of moving window, namely piecewise OSC (POSC), to pretreat the spectra matrix and eliminate the local variance, thus the spectra matrix pretreated was taken as the new independent variables matrix, then the PLS algorithm was applied to build the calibration model. Finally, application of the proposed POSC-PLS approach to the EAS quantitative analysis of the polyaromatic hydrocarbons (PAHs)was presented for comparison with the MLR (multiple linear regression), PLS and OSC-PLS methods. The result indicates that the POSC-PLS approach by performing POSC prior to calibration not only can improve the model accuracy, but also decreases the PLS factors compared to the models obtained by the above rest methods and so its resulting model becomes more concise. The removal of orthogonal components from the response matrix is greatly facilitated simply by considering localized spectral features. So, preprocessing with POSC was shown to benefit the multivariate PLS model because it performed a localized regression modeling procedure that differs from that of PLS. At the same time, the POSC is a potential chemometric technique in the pretreatment of various spectra.