Elliptical novelty grouping for on-line short-turn detection of excited running rotors

A technique for the detection of shorted turns in the field-windings of operating synchronous turbine-generators is described. The measuring method used is the twin-signal sensing method, where pulses are injected into each terminal of the rotor. The reflected signals are subtracted to produce signature signals that contains information about the rotor's state. The signature signals are sampled and accepted or rejected as valid based on an outlier detection criteria. Novelty detection is applied to the accepted signals. The current signature signal is compared to a range of signature signals taken when the rotor was known to be free of shorts. If the current signature signal strays too far from the known healthy signal, the signature signal is 'novel' and the possibility of a short is declared. The method was tested on a running test rotor with voltage excitation. A comparison of methods shows that an elliptical novelty grouping algorithm gives highly accurate results.