A supervised learning procedure for classification of steel samples analyzed by means of optical emission spectroscopy was developed. The method should be applicable in simple portable spectrometers, for various groups of materials and working relatively fast. Data vectors extracted from digital spectra of unknown samples are compared with average vectors evaluated from the data vectors of repeated measurements of reference samples. The classification is carried out on the basis of the multivariate distance between the data vector of the unknown sample and the nearest average reference vector and its deviation. The supervised learning procedure was tested by 10 steel samples which could be successfully classified.