Rolling element bearing fault detection using density-based clustering

Fault detection is a critical task in condition-based maintenance of rolling element bearings. In many applications unsupervised learning techniques are preferred in fault detection due to the lack of training data. Unsupervised learning techniques such as k-means clustering are most widely used in machinery health monitoring. These methods face two challenges: firstly, they cannot cluster non-convex data, which may have arbitrary shape; secondly, no rule has been established for these techniques to find a fault threshold. This paper introduces a fault detection methodology based on density clustering to address these challenges. This methodology assumes that data from healthy bearings is located in regions with a high density and data from faulty bearings is located in low density regions. By finding boundaries of these regions, which may be non-convex, data from faulty bearings can be identified. In this paper the value of the density for healthy bearings and faulty bearings is evaluated. The rate of change of the density from healthy to faulty is identified as a fault threshold. The methodology is validated by experimental data. This methodology can be applied to applications where faulty data are too difficult or costly to acquire. Also it can be used in applications where fault thresholds are difficult to determine.