Research On Model Identification Method Based On Spectral Information

The near infrared spectral analysis technology is a kind of efficient, no damage to sample and indirect measurement analysis technology, it has been widely used in petroleum chemical industry, food, pharmaceutical and other industries. The key of the spectrum is to establish an appropriate mathematical model with measured sample spectral information and sample target content standard measurements or category, then the model been used to predict unknown samples. The main mathematical tools of spectral analysis is chemometrics which include the sample of calibration and validation set selection, data preprocessing, quantitative and qualitative modeling and main method of model transfer, etc. Based on the above method of chemometrics, this paper respectively researched on calibrating validation set selection and qualitative modeling method.The sample selection of calibration set and validation set correction is very important to spectral multivariate analysis. Due to restrictions of the actual factors, however, most of the samples collected in the interval of properties distribution is not uneven. This paper explore the effect of uneven samples for modeling results, and proposed a new method on the current method which is Rank-KS. The method can comprehensive considerate in spectrum space and properties space to improve the uniformity. Application results have verified that the best prediction was achieved by using Rank-KS. Especially, for the distribution of sample set such as more in the middle and less on the boundaries, or none in the local, prediction of the model constructed by calibration set selected using Rank-KS can be improved obviously.Qualitative analysis technology is an important field of near infrared spectral technology, it is only need to judge the categories of unknown samples but not the content or information of components in many cases, namely the near infrared spectrum pattern recognition, and one of the most important is feature extraction to the original spectral data for pattern recognition. This paper analyzed the most commonly used feature extraction method of near infrared spectral-Principal Component Analysis (PC A), and a new feature extraction method is introduced which named "KL transform based on the supervision pattern recognition", it takes advantage of sample classification information to extracte the features which include centralized in class mean vector and centralized feature vector in the classification of information extraction. The qualitative model was established by SIMCA and LS-SVM based on PC A feature extraction, and compared the proposed optimal KL transform, the experimental results show that the validation set samples all correct classification by the method proposed in this paper. So in the near infrared spectrum identification model, the KL transform based on the supervision pattern recognition is a simple and effective method for NIR qualitative analysis.