Libs Feature Variable Extraction Method Based on Convolutional Neural Network

Abstract

The extraction of feature variables can significantly reduce the dimension, simplify the features, and improve the accuracy of quantitative models, which is of great significance in spectral data preprocessing using laser-induced breakdown spectroscopy (LIBS). A feature variable extraction method based on convolutional neural network was proposed. The LIBS spectral data was subjected to multilayer two-dimensional convolution through the series connection of the convolution structure and an InceptionV2 module, and the multi-level features were gradually extracted to find the optimal feature variable combination. Finally, the prediction results were obtained by the fully connected layer. In order to verify the applicability of the extracted features, the extraction results of the convolutional neural network were directly input into random forest to construct a quantitative model. In this paper, the concentration of K element in the mixed solution prepared by the laboratory was tested. The determination coefficients R² of the CNN–RF training set and the test set reached 0.993 and 0.990, respectively. The root mean square error (RMSEC) of the training set and the root mean square error (RMSEP) of the test set were 0.0067 and 0.0084 wt.%, respectively, and the average relative error reached 8.533%. The evaluation parameters were significantly better than the extraction results of least absolute shrinkage and selection operator, principal component analysis and SelectKBest. The results show that the convolutional neural network can effectively extract LIBS characteristic spectral lines and improve the accuracy of quantitative analysis.