Early Detection of Bearing Damage by Means of Decision Trees

This paper presents a procedure for early detection of rolling bearing damages on the basis of vibration measurements. First, an envelope analysis is performed on bandpass filtered signals. For each frequency range, a feature indicator is defined as sum of spectral lines. These features are passed through a principal component model to generate a single variable which allows to track change in the bearing health. Thresholds and rules for early detection are learned thanks to decision trees. Experimental results demonstrate that this procedure enables early detection of bearing defects.