Missing Data Recovery And Alternating Residual Trilinearization Applied To Multivariate Calibration And Calibration Transfer

The research work in this thesis focuses on the multivariate calibration and prediction with deformity data.The existence of the Rayleigh scattering tends to increase the difficulty of the fluorescence spectral analysis. In order to eliminate or reduce the interference of Rayleigh scattering, a combination of missing data recovery iterative algorithm and principal component analysis (MDR-PCA) is developed. In the case of single component excitation-emission matrix (EEM) fluorescence spectra, the Rayleigh scattering block is recovered by setting the appropriate data to missing area during the PCA iteration progress. The MDR-PCA method accurately extracts the pure spectra of the component and reconstructs EEM spectra existing scattering region. For multi-component EEM fluorescent spectra, the MDR-PCA method with the constraint that forcing the signal values in the missing area monotonously decline can effectively correct scattering region area after choosing the appropriate number of principal components. In addition, the parallel factor analysis (PARAFAC) coupled with MDR method is also examined to eliminate the impact of Rayleigh scattering signal. In the process designed, the MDR follows the iterative calculation of the trilinear decomposition and repairs the missing data meanwhile. Results show that the analytical solutions obtained by MDR-PARAFAC method in the simulation data, phenolic fluorescence and molecular beacon fluorescence data sets are highly consistent with the true values. It proves that MDR-PARAFAC method can make full use of effective information of EEM spectra and effectively avoid the impact of interference signals.Molecular beacon (MB) probe technology has extensive applications in the analysis of small molecule, protein and nucleic acid. But the very few kinds of labeled fluorophores on MB and the wide bandwidth of their fluorescence both limit its application in multicomponent simultaneous detection. In order to improve the detection performance of the MB analysis, we provide a new second-order calibration method, alternating residuals trilinearization (ART) algorithm, which could solve the spectrum overlapping problem in multicomponent fluorescent MBs and simultaneously determine multiple targets. Compared to the conventional second-order calibration method PARAFAC, the number of components could be obtained automatically in the convergence process in ART algorithm, rather than be determined individually prior to the analysis. Besides, the missing data recovery algorithm is coupled with the ART method to avoid the impact of Rayleigh scattering signal interference. Though the excitation-emission spectra of the three MBs are heavily overlapped, it also can be resolved by the algorithms well. The satisfactory quantitative results and resolved spectra are obtained.To keep the multivariate calibration model built on one instrument usable for other instruments is indispensable for the successful application of the NIR analysis methods. Missing data recovery algorithm is applied for the transformation of the spectra obtained from various instruments to the direction of the destination instrument. In this method, the spectra of several standardization samples from the destination instrument and the source instruments are combined to construct the transformation matrix. Then the sample spectra from the source instrument are attached to the matrix and the corresponding spectra area for the destination instrument is set as missing block. After MDR-PCA processing, the recovered data blocks are designated to the transformed spectra. The prediction performances of MDR method and global model, which builds models by mixing the calibration spectra from all instruments, are compared. When there are only 2 instruments, the two methods get similar RMSEP improvements. When the number of instruments is 3, MDR has better prediction accuracy than global models. Furthermore, global modeling needs the collection of spectra of the calibration set on every instrument. Meanwhile MDR only need limited standardization samples (usually less than 10). MDR method is more preferable in performance and efficiency than global modeling methodIn this contribution, a novel multivariate calibration model maintenance method named SST is proposed. The performance of SST is evaluated and compared with three other methods (i.e. global PLS, slope and bias correction (SBC) and piecewise direct standardization (PDS) designed for the same purpose), using two spectral data sets. For both data sets, SST achieved the best results. Consequently, SST is a very competitive calibration model maintenance method that should have wide applicability in on-line/in-line process monitoring.On the basis of the previous work above, the SST and MDR method are applied for the maintenance of near-infrared analysis model for tobacco chemical composition monitoring. As well, two kinds of spectral data pretreatment methods (the first order derivative method and discrete wavelet transform method) are also discussed. The results demonstrate that the implementation of SST is rather simple and can effectively correct the spectral variations induced by changes in instrument. Only one model parameter needs to be optimized in SST, and the performance of model maintenance is significantly improved. Compared to the three other methods, SST has more obvious advantage when the difference between the instruments is large or complex. Although the result obtained by MDR method is comparable to SST, the efficiency of MDR is lower than SST because every transferring of spectrum in MDR requires NIPALS iterative calculation.