Radkowski S., Dybala J., Gontarz S.

The problem of optimal choice of most informative diagnostic features of signal, particularly the vibroacoustic signal, is presented. Taking into account that the data are in form of available attribute vectors, many methods for discriming between two classes have been developed. In this paper we review own and literature results that will be relevant to the solution of diagnostic inference. The special attention will be paid on the method of geometrical features selection and blind source separation. The first method presented here use two criteria of ability to separate (isolate) classes of an object’s state: the criterion of average scatters and the original criterion of number of prototypes of classes. Method can be successfully used for initial analysis of input data to a neural network dealing with recognition of patterns of an object’s state and classification of an object’s state. Concerning blind separation, the method which uses algorithm of blind equalization (BE) by iterative application of different lengths of equalizers is presented. This approach allows to estimate subsignals in various frequency bands. Main features of this solution is possibility of discovery and identification diagnostically useful information even that, when it is hidden by relative larger noise and interferences.