Vibrodiagnosis helps in detecting incipient faults in rotating machines like pumps and generators. Early detection prevents from undesired breakdown of the machine and allows to schedule maintenance times. The application of neural networks in classification of the rotating machine condition has been described in this work. Different types of networks and methods on feature extraction was described and compared. Additionally it was proposed a novelty feature set consisted of harmonics from vibration spectrum. The set were combined with using of probabilistic neural networks which has been modified that it could recognize defects that did not occur in the training set. Such architecture was tested in detection of two defects, shaft misalignment and mass unbalance. It was found that such network works better than a multi layered perceptron with statistical features.