New Methods For The Study Of Near Infrared Spectrum Analysis Based On Signal Processing

Near infrared (NIR) spectroscopy is an analytical method with the advantages of great efficiency, high speed, non-destructivity, environment friendly, and low costs. It has been widely used in different fields. Near infrared technique needs the help of chemometrics methods to establish the model to obtain a better quantitative analysis performance due to its defect of wide peaks and overlapping. During modeling process, spectral preprocessing and modeling are very important jobs. In the dissertation, corrected method for interferogram, multivariate calibration methods based on the mixed model of samples, detection method for abnormal spectrum based on the mixed model of samples, and minimization method of relative error were investigated for the quantitative analysis of NIR spectrum. The main research contents of the dissertation are as follows:1. Interferogram is the original measured signal of the Fourier transform spectrometer. The phase error of interferogram will affect the quality of the spectrum. On the basis of origins analysis of the phase errors of interferograms, a new model is proposed for the phase correction of Fourier transform spectrometer interferograms. The new model is composed of three parts including digital all-pass filter, delayer and phaser. Therefore, the new model is named ADP model. When the ADP model is applied to phase correction, the particle swarm optimization (PSO) algorithm is applied to design of digital all-pass filter. Simulation experiments of phase correction of double-sided interferograms using ADP model based on PSO algorithm are given. Simulation results show that it is easy to implement higher phase correction precision by using ADP model based on PSO algorithm with less iteration number. Application experiment shows that, the predict error of spectrum corrected by ADP model based on PSO algorithm is less than those of spectrum corrected by Mertz method.2. A novel algorithm for linear multivariate calibration that can generate good prediction results is proposed. This is accomplished by the idea of that testing samples are mixed by the calibration samples in proper proportion. The algorithm is based on the mixed model of samples and is therefore called MMS algorithm. With both theoretical support and analysis of two data sets, it is demonstrated that MMS algorithm produces lower prediction errors than partial least squares (PLS2) model, has similar prediction performance to PLS1. At the condition of the lack of some component information, MMS algorithm shows better robustness than PLS2. In the anti-interference test of background and noise, MMS algorithm performs better than PLS2.3. A blind identification method of abnormal spectrum based on the mixed model of samples (MMS) is proposed. There,’blind identification’means that the method can identify the abnormal spectra without known the prediction content value. The method is fit for the prediction samples whose contents are unknown. For the prediction sample, there are three kinds of abnormal spectrum. The origins of abnormal spectrum include measurement background change, instrumental noise increase, and prediction samples containing non-calibration content. The identified processes can be described as follows. Firstly, mixed vector of prediction sample is calculated according to the mixed model of samples. Secondly, estimated spectrum of prediction sample is calculated according to the mixed ratio and the spectrum of calibration samples. Thirdly, the difference between estimated spectrum and the measuring spectrum is calculated. Lastly F-Statistical test is carried out to identify the abnormal spectrum according to the variance. The method is compared with the MMS, PLS, and the root mean square error in spectral residual (RMSSR) algorithms. In the experiment, it is assumed that the contents are unknown for this method. For MMS and PLS, the contents are known, and when the prediction error is larger than three times the root mean square error of prediction (RMSEP), the spectrum is identified as abnormal spectrum. Experimental results show that this method has better recognition performances for abnormal spectrum caused by measurement background change, instrumental noise increase, and the condition of prediction samples containing non-calibration sample than MMS, PLS and RMSSR algorithm. This method provides a new approach to detect the spectrometer performance including the background change and noise increase in advance.4. RMSEP is usually used as the evaluation criterion for most of multivariate calibration methods. According, root-mean-square relative error of prediction (RMSREP) is seldom considered during the process of multivariate calibration. A new method addressed the problem of minimizing prediction relative error is proposed for multivariate calibration. The method is based on the use of back-propagation artificial neural network (BP-ANN). The regression objective of the new method is to minimize the prediction relative error by changing the output values of BP-ANN. The method consists of basic algorithm and enhanced algorithm. Experiments show that basic algorithm produced lower prediction relative error than PLS method. Considering the value of RMSEP, basic method and PLS method have a similar prediction performance. For the enhanced algorithm, the basic algorithm and PLS are mixed to minimize the RMSEP and RMSREP simultaneously.