Research Of Semi-Supervised Online Sequential Extreme Learning Machine Regression Algorithm Based On NIR Spectral

The Near infrared (NIR) Spectrum monitors samples fast and conveniently, with low cost and without destroying the samples, which is widely used in the food industry, agricultural production, drug manufacturing and other areas. In the applications of online measurement of NIR spectra, it is difficult and expensive to get the samples’physicochemical indexes and other labels, mostly only a small amount of NIR spectroscopy samples have corresponding labels, while a large amount of samples don’t have labels. The traditional NIR analytical methods usually train the relationship between the spectrum and the label in supervised batch mode, which makes the model poor generalization ability, and training time high complexity. Extreme Learning Machine (Extreme Learning Machine, ELM) can fast and effectively train high dimension, nonlinear NIR data and obtain network structure of good generalization. Therefore, based on ELM algorithm, this thesis puts forward a semi-supervised online sequence ELM regression algorithm (SSOSELMR), which is used to solve the problem that a small number of labeled NIR spectrum semi-supervised modeling online. The main work in this thesis as follow:Semi-supervised ELM regression model is established. Based on only a small amount of NIR spectroscopy samples with labels, a large amount of samples don’t have labels NIR data, Give the modeling method and the derivation process of semi-supervised ELM, estimate the label confidence to select credible unlabeled samples, and use the co-training of ELM to label reliable samples, then use the new labeled samples to calculate the output weights of semi-supervised ELM regression.Based on the online sequential characteristics of the semi-labeled NIR data, further a semi-supervised online sequence ELM regression model is established. Give the modeling method and the derivation process of semi-supervised online sequence ELM, on the basis of Online sequence Extreme Learning Machine, set the confidence level evaluation function to select unlabeled reliable samples, use the co-training of OSELM to label reliable sample, then calculate the output weights of semi-supervised Online sequence ELM regression.In this thesis, several groups of NIR data are used to verify SSOSELMR algorithm. Experimental results show that under different label rate, compared with OSELMR, training time of SSOSELMR have be improved, compared with COREG, test error of SSOSELMR have be improved, which illustrates the SSOSELMR can set up semi-supervised online learning model with semi-labeled NIR data.