Feature extraction for novelty detection as applied to fault detection in machinery

Novelty detection is a pattern recognition technique used when there is one well characterized normal state and the abnormal (or novel) states are poorly described because of lack of data. Data deficiency of these states may arise due to cost and difficulty in measuring them – e.g. failed equipment states in equipment health monitoring. Normal pattern recognition techniques have a wide array of methods for reducing the number of features initially employed to characterize classes. These techniques are of limited use in novelty detection primarily because they are focused on representing the data accurately in a subspace rather than on finding a subspace where the classes can easily be discriminated, or they are optimized to distinguish between all classes rather than on focusing on distinguishing solely between normal and abnormal classes. The proposed methodology will enable feature extraction in unbalanced classification problems where the well-sampled normal data is expected to be orbited by the under-sampled fault data. The technique will be demonstrated to work well with artificial and actual machinery health data.

Research highlights
► Proposed a simple feature reduction technique optimized for novelty detection.
► The technique focuses on optimal discrimination in the context of novelty detection.
► An improvement is suggested to the technique’s dependence on outlier data.
► The approach was compared with those existing in the literature and was found to provide competitive results.