Rotor Faults Diagnosis Using Feature Selection and Nearest Neighbors Rule: Application to a Turbogenerator

Among failures that are observed in power plants, rotor faults have often been recorded. Thus, radial flux probes have been introduced in most generators to anticipate heavy rotor faults such as rotor ground faults. However, the commonly used signal processing done with those rotor flux measurements makes the diagnosis complicated. The aim of this paper is to present a diagnosis method based on statistical pattern recognition using flux probe and classical electric measurements. For that aim, a specific experimental setup has been designed to perform the methodology. This experimental setup is a small-scale prototype of a nuclear plant generator, which is, in fact, a direct-current-excited synchronous machine. In this generator, electrical and mechanical rotor faults can be carried out. Sixteen functional states have been performed for five operating points. From each measurement, a list of scalar parameters is extracted. Then, to reduce their number, a selection stage is achieved through the Fisher criterion and the sequential backward selection algorithm. Finally, the classification stage is performed using the k-nearest neighbors rule with Euclidian and Mahalanobis distances. As a result, the methodology developed removes diagnosis ambiguities of the commonly used signal processing by clearly splitting different types of faults.