Abstract:

Anomaly detection is a critical task in condition-based maintenance of machinery. In many applications clustering-based anomaly detection is preferred due to its ability to analyze data which may not follow a well studied distribution and are unlabeled. This paper introduces a density-based clustering method for machinery anomaly detection. This method assumes that the data from healthy states are located in regions with high densities and the data from faulty states are located in low density regions. By finding the boundaries of these regions, data from the anomalous states can be identified. The values of the densities for healthy machinery and faulty machinery are evaluated. The rate of change of the density from healthy to faulty is identified as a fault threshold. This method can be valuable for applications where faulty data are too difficult or costly to acquire.

This article is available to CALCE Consortium Members for personal review.