Optimum Feature Extraction and Selection for Automatic Fault Diagnosis of Reluctance Motors

An intelligent approach artificial neural network (ANN) combined with genetic approach (GA) is presented for detection of stator winding related fault of switched reluctance machine. Switched reluctance machine (SRM) is known to be fault tolerant, however, is not fault free, and questions emerge as to powerful diagnostic methods. This paper takes an in-depth look at winding open-circuits 'the worst case' in this particular machine. Various cases are considered, falling in two distinct categories: (i) when an entire phase is opened; (ii) when only part of a winding is opened. Therefore, application of classification method is very necessary to get the exact information to classify and to obtain a more complete labeling, and so, a more powerful diagnosis. An appropriate features extraction and features selection techniques should be incorporated. In this proposed method, smoothing Time-Frequency Representation (TFR) from a time-frequency ambiguity plane is used to extract features from torque time signals. In order to reduce the number of the features, a GA is suggested to select optimal ones. The new features provide more sensitive information for a classifier. The proposed features feed a simple non-linear classifier based ANN which separates almost perfectly between normal and faulty conditions, with also very high diagnostic accuracy between the faulty classes. Experiments are carried out using a laboratory apparatus to show the how various configurations of the system are able to detect different classes of faults. The proposed method successfully distinguished the difference, and classified SRM open-circuit faults correctly.