Search

To improve the accuracy of laser-induced breakdown spectroscopy (LIBS) in the classification of tea leaves, a modified adaptive mutation probability for genetic algorithm (GA) was proposed to optimize support vector machines (SVM). The implementation process of the GA-SVM algorithm was discussed, and the key parameters were analyzed. The penalty factor and kernel function parameters in SVM were optimized by GA. The spectral line intensities of Mg (279.55 nm), Mn (279.83 nm), Mg (280.27 nm), CN (0–0) (388.34 nm), Ca (393.37 nm), Al (396.15 nm), Ca (396.84 nm), C2 (0–0) (516.45 nm), Fe (517.46 nm) and K (766.49 nm) compared to C (247.86 nm) were selected as the analysis indexes according to the differences of LIBS spectra. LIBS spectra pre-processed by multiple scattering correction (MSC) with the optimal input feature were used to construct the GA-SVM model for different tea species. The results showed that the average correct classification rate was 99.73% in the training set. In addition, the average accuracy was 98.40% in the test set. The classification accuracy of the improved GA-optimized SVM was obviously higher than that of cross validation-support vector machine (CV-SVM) and particle swarm optimization-support vector machine (PSO-SVM). This work demonstrates that the GA-SVM can avoid the blindness of parameter selection and can effectively increase the accuracy of tea classification. The results indicated that LIBS is an effective technology in identificating tea leaves; moreover, it has a real-time, rapid and reliable measurement prospect. [ABSTRACT FROM AUTHOR]

In order to improve the stability and prediction accuracy of a Laser Induced Breakdown Spectroscopy (LIBS) quantitative analysis model a chemometrics method named PSO-SVM (particle swarm optimization-support vector machine) is proposed for the quantitative prediction of heavy metal chromium (Cr) in pork by analyzing the spectral data obtained by LIBS. The usable spectral region from 424.77 to 429.99 nm was identified. The normalized preprocessing of spectra was employed. Also, the spectral dimension was reduced through principal component analysis (PCA). The PSO algorithm was applied to select the SVM hyper-parameters including the regularization and kernel parameters. The calibration model of heavy metal Cr in pork was established using the optimum hyper-parameters obtained by PSO-SVM. The performance of this PSO-SVM model was compared with a synergy interval partial least squares (siPLS) model and a multilayer perceptron artificial neural network (MLP-ANN). The feasibility of these three methods was examined on the test sample set. The experimental results showed that the proposed method (PSO-SVM) achieved better prediction accuracy robustness compared with the other two quantitative analysis methods. The calibration results of siPLS and MLP-ANN models are similar, and the fitting degree of the siPLS modeling effect is not high, while the MLP-ANN verification results are not ideal. For the PSO-SVM model, the correlation coefficients (R2) of the calibration and validation sets were higher than 0.99. The mean absolute percentage error (MAPE) of the calibration set is 0.0792%, and the MAPE of the validation set is 1.9395%. The RMSE values of the calibration and validation sets were 0.0245% and 0.6961%, respectively, which were much lower than the other two models. This investigation displayed that the PSO-SVM algorithm exhibited excellent model robustness and prediction accuracy. There is definite practical significance and application value in the monitoring of meat safety by analyzing LIBS data coupled with PSO-SVM methods. [ABSTRACT FROM AUTHOR]