INDEPENDENT COMPONENT ANALYSIS; SUPPORT VECTOR MACHINES; NEURAL-NETWORK WNN; QUANTITATIVE-ANALYSIS; RANDOM FOREST; FLY-ASH; WAVELET-TRANSFORM; UNBURNED CARBON; SIGNAL; MODEL

The classification and identification of coal ash contributes to recycling and reuse of metallurgical waste. This work explores the combination of the laser-induced breakdown spectroscopy (LIBS) technique and independent component analysis-wavelet neural network (ICA-WNN) for the classification analysis of coal ash. A series of coal ash samples were compressed into pellets and prepared for LIBS measurements. At first, principal component analysis (PCA) was used to identify and remove abnormal spectra in order to optimize the training set for the WNN model. And then, ICA was employed to select and optimize input variables for the WNN model. The classification of coal ash was carried out by using the WNN model with optimized model parameters (the number of hidden neurons (NHN), the number of iterations (NI), the learning rate (LR) and the momentum) and input variables optimized by ICA. Under the optimized WNN model parameters, the coal ash samples for test sets were identified and classified by using WNN and artificial neural network (ANN) models, and the WNN model shows a better classification performance. It was confirmed that the LIBS technique coupled with the WNN method is a promising approach to achieve the online analysis and process control of the coal industry.

Wines; LIBS; Classification; Principal component analysis; Random forest; Matrix effect;LIQUID-LIQUID MICROEXTRACTION; ICP-MS; GEOGRAPHICAL ORIGIN; PROTECTED DESIGNATION; MULTIELEMENT ANALYSIS; SINGLE-PULSE; IDENTIFICATION; SAMPLES; LIBS; METALS

Laser-induced breakdown spectroscopy (LIBS) has been applied to classify French wines according to their production regions. The use of the surface-assisted (or surface-enhanced) sample preparation method enabled a sub-ppm limit of detection (LOD), which led to the detection and identification of at least 22 metal and nonmetal elements in a typical wine sample including majors, minors and traces. An ensemble of 29 bottles of French wines, either red or white wines, from five production regions, Alsace, Bourgogne, Beaujolais, Bordeaux and Languedoc, was analyzed together with a wine from California, considered as an outlier. A non-supervised classification model based on principal component analysis (PCA) was first developed for the classification. The results showed a limited separation power of the model, which however allowed, in a step by step approach, to understand the physical reasons behind each step of sample separation and especially to observe the influence of the matrix effect in the sample classification. A supervised classification model was then developed based on random forest (RF), which is in addition a nonlinear algorithm. The obtained classification results were satisfactory with, when the parameters of the model were optimized, a classification accuracy of 100% for the tested samples. We especially discuss in the paper, the effect of spectrum normalization with an internal reference, the choice of input variables for the classification models and the optimization of parameters for the developed classification models. (C) 2017 Elsevier B.V. All rights reserved.