Geometrical method of selection of features of diagnostic signals

The paper presents the method of selection of features of diagnostic signals relying on the geometry of observation space. The method presented here uses 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. The method can be successfully used for initial analysis of input data to a neural network dealing with recognition of patterns and classification of an object's state. It also possible to use the algorithm presented here for determining the prototypes of classes needed for creating the training set for the neural network.