The paper deals with acquisition of diagnostic knowledge that is relevant for detection and isolation of a special class of malfunctions of rotating machinery called "shaft misalignment". To detect a misalignment of the given shaft supported by multiple journal bearings, decision trees have been applied. These trees have been discovered in a database collected in a numerical experiment performed by the well-verified simulation system. A novel approach to definition of classes of misalignment has been introduced. Several new methods of selection of attributes and evaluation of classifier's performance have been suggested and verified. Finally a new method of diagnosing misalignment of rotating machinery has been formulated. This method may be efficiently implemented for real-existing rotating machinery.