A Distance Metric Learning Based Health Indicator for Health Prognostics of Bearings

The accuracy of bearing health prognostics highly depends on the constructed health indicator. This paper presents a health indicator construction method based on distance metric learning (DML). First, multiple features are extracted from the raw monitoring vibration signals, including a designed feature named as average energy of fault frequency band (AEFFB). Then, the optimal features are selected from the extracted features according to monotonicity and correlation metrics, where the correlation metric is calculated by the determination coefficient in random forest regression. Finally, a distance metric is learned utilizing DML, which is then used to construct a health indicator with the help of self-organizing map. The effectiveness of the proposed health indicator is validated by accelerated degradation data of bearings.