Spectral Abnormality Detection And Qualitative Analysis Of Petroleum Products Based On Local Approaches

In the methods of oil classification, spectral analysis gets more and more attention due to its rapid measurement, nondestruction to samples, high efficiency and low cost.This thesis aims to distinguish kinds of petroleum products and classify product gasoline by source and brand based on NIR spectroscopy and Raman spectroscopy. Local approach is used in spectral outlier detection and recovery, model parameter optimization and classification algorithm modification. The detailed research work is given as follows:1. An improved algorithm based on local approach to remove cosmic spikes in Raman Spectra for online monitoring is proposed. In this algorithm, a new scheme composed of intensity identification and local moving window correlation analysis is introduced for cosmic spike detection; intensity identification based on derivative spectra and local linear fitting approximation are used for the recovery of cosmic spikes. The algorithm is proved to be simple and effective, which has been applied in an online Raman instrument.2. A new classification method based on NIR spectroscopy for gasoline brand recognition is proposed. In this method, Isomap(a manifold learning algorithm) is used to reduce dimensionality of spectra, which is different with traditional dimensionality reduction by PCA, and K-nearest neighbor (KNN) algorithm is implemented after dimensionality reduction. Therefore, this method is called the Isomap-KNN algorithm. The classification experiment results show that the manifold learning algorithm can obtain more feature information of gasoline brand than PCA during dimensionality reduction.3. A novel local weighted LSSVM algorithm based on Raman spectroscopy is proposed to classify gasoline samples by source and brand. The weight is constructed based on correlation coefficient R and this algorithm can be denoted as R-weighted LSSVM. In this algorithm, both of Euclidean distance and correlation coefficient are considered to select neighboring samples. Local approach is realized by the RBF kernel and the weight. LDA based on PCA, LSSVM, local LSSVM and R-weighted LSSVM are compared in the classification experiment. Experimental results show that Raman spectroscopy is an effective means to classify gasoline brand and origin, and the R-weighted LSSVM algorithm gives the best classification result.4. A simple method based on correlation analysis for the classification of petroleum products is proposed. This method gives good classification results because of significant differences between Raman spectra of petroleum products. This method costs little calculation time and human interference. Moreover, it can be easily implemented in the practical application.5. A simple method used to optimise the parameters of LSSVM is proposed. This method optimises the LSSVM model by adjusting the number of neighboring samples instead of adjusting the RBF kernel parameter. The range of adjustment is also reduced. The traditional punishment coefficient y is replaced by characteristic punishment coefficient which can be easily set in a wide range. Experiment results show that this method has similar optimization performance with other complex optimization methods.