CLASSIFICATION OF ENERGY DISPERSION X-RAY SPECTRA OF MINERALOGICAL SAMPLES BY ARTIFICIAL NEURAL NETWORKS

摘要

Automatic classification of different mineralogical samples into 12 prespecified classes using Kohonen artificial neural networks (ANNs) is studied in comparison with standard chemometric techniques: hierarchical clustering and principal component analysis. The mineral types into one of which the unknown samples should be classified are pyrrhotite, pyrite, chalcopyrite, pentlandite, magnetite, biotite, albite, talc, chlorite, lizardite, dolomite, and amphibole. The basis for classification are 15-dimensional EDX spectra of individual grains taken from large matrices of compositions each containing a variety of grains belonging either to the same or to different minerals. The discussed classification procedure is based on the 15-15-15 Kohonen neural network cube. The classification results are displayed on the 15 x 15 Kohonen top-map. From the 15 weight levels of the 15-15-15 Kohonen ANN 12 logical rules that allow one to classify unknown samples into one of 12 classes are extracted. The 100% correct classification of samples using the suggested 12 logical rules is enabled by only seven of out of 15 intensity lines from each (energy-dispersive X-ray) EDX spectrum. It is shown that the Kohonen ANN allows one to draw conclusions and logical rules based on the weight patterns formed during the training in the ANN. [References: 14]