Abstract
The paper presents the use of inductive machine learning for selecting appropriate features capable of detecting washing machines that have mechanical defects or that are wrongly assembled in the production line. The input data are vibration signals from different points of washing machine's surface when it operates for a couple of minutes in centrifuge mode. The signal is normalized and transformed to the frequency domain using Fast Fourier Transforms (FFT). An adequate example database is constructed from samples of different machine status. Our approach lends basic concepts contained in the algorithmic family of ID3 and C4.5. Certain parts of the algorithmic process contained in the above machine learning tools are used in order to construct a feature selection methodology based on information entropy criteria. The selected features are then used by specific classification techniques for achieving successful discrimination among different types of fault and normal operation. The overall methodology, using real data acquired within the European Community funded project called MEDEA, obtains a high rate of fault detection exceeding 87% of successful classification.