The fault diagnosis field is in a continuous movement towards the generation of more reliable and powerful machine health monitoring schemes. Improved data processing methodologies are required to reach high diagnosis demands. For that reason, a contribution in motor fault classification methodology is presented. Different physical magnitudes such as phase currents, voltages and vibrations, are acquired from an electromechanical system based on Brushless DC motor. Statistical features, from time and frequency domains, are calculated to supply a classification algorithm based on Neural Network and enhanced by Genetic Algorithm. The significance of feature space dimensionality, related with the number of used features, for classification success is analyzed. The combination of a feature selection technique (by Sequential Floating Forward Selection), with a feature extraction technique (by Principal Component Analysis), is proposed as a novel hybrid feature reduction methodology to improve the classification performance in electrical machine fault diagnosis. The proposed methodology is validated experimentally and compared with classical feature reduction strategies.

Peer Reviewed