Abstract

Three main classification methods, kNN, Bayesian, and C4.5 inductive learning were investigated for their ability to differentiate types of either aluminium or steel alloys. Simulation experiments were performed with two generated test data sets to evaluate their potential for classifying samples based on elemental compositions. The first test set had normally distributed errors of 25% RSD. This data set simulated results of a preliminary scan of semi-quantitative accuracy to be used for selection of a calibration methodology. The second test set had 5% RSD and simulated results of a quantitative analysis, where the identity of the sample was to be determined (identification). The Bayesian method provided the best results with prediction accuracy rates of 89.8% for the first goal (methodology selection) and 100% for the second (identification) in Al alloys. Steel alloys were also classified best with the Bayesian method, with 96.7% and 100% accuracy, respectively. A detailed flow chart for sample recognition is proposed, which optimizes both prediction accuracy and calculation speed.