Abstract:In near infrared spectral quantitative analysis, the precision of measured samples' chemical values is the theoretical limit of those of quantitative analysis with mathematical models. However, the number of samples that can obtain accurately their chemical values is few. Many models exclude the amount of samples without chemical values, and consider only these samples with chemical values when modeling sample compositions' contents. To address this problem, a semi-supervised LS-SVR (S2 LS-SVR) model is proposed on the basis of LS-SVR, which can utilize samples without chemical values as well as those with chemical values. Similar to the LS-SVR, to train this model is equivalent to solving a linear system. Finally, the samples of flue-cured tobacco were taken as experimental material, and corresponding quantitative analysis models were constructed for four sample compositions' content(total sugar, reducing sugar, total nitrogen and nicotine) with PLS regression, LS-SVR and S2 LS-SVR. For the S2 LS-SVR model, the average relative errors between actual values and predicted ones for the four sample compositions' contents are 6.62%, 7.56%, 6.11% and 8.20%, respectively, and the correlation coefficients are 0.974 1, 0.973 3, 0.923 0 and 0.948 6, respectively. Experimental results show the S2 LS-SVR model outperforms the other two, which verifies the feasibility and efficiency of the S2 LS-SVR model.

A Novel Approach to Nir Spectral Quantitative Analysis : Semi-Supervised Least-Squares Support Vector Regression Machine