On feature extraction for condition monitoring using time series analysis and distance techniques

The paper comprises two parts which explore some possibilities for feature extraction and classification in condition monitoring of robot joints from their measured acceleration signatures. The symmetrized Itakura distance is used to form features and to develop several classifiers to distinguish between signals coming from joints with different amounts of backlash. The classifiers are tested and their performance is discussed and compared. In the second part some nonlinear dynamics characteristics are extracted directly from the measured transients and considered as possible features for defect detection and classification. Finally, the potential of nonlinear dynamics for robot joint dynamics modeling is highlighted.