ReliefF-Based Feature Ranking and Feature Selection for Monitoring Induction Motors

A feature is a measured property of a monitored system. Feature extraction in condition monitoring requires domain knowledge about a system and its possible fault cases. To find the most sensitive features for fault patterns, one has to evaluate the relevancy of features. In this paper the authors use ReliefF, which is a K-nearest neighbors-based features selection algorithm, to evaluate the extracted features from an induction motor dataset. The dataset contains data from nine different health states of an induction motor. Feature relevancy is calculated for each health state. The selected features are fed into a simple Bayesian binary classifier to calculate the most likely health state. The method provides insight into the relevance of features by sensor type and also by signal processing type. The evaluation of similarity among the selected features can help identify similar faults. The results obtained emphasize the importance of domain knowledge in proper design of features. Furthermore, by considering experimental data obtained for multiple loading and noise conditions, the feature selection method indicates features which are best suited for diagnosing specific faults, regardless of external conditions. Such information can support the creation of robust monitoring systems.