Supervised Discretization of Multivariate Sensor Data for Fault Detection in Bearing Shaft Systems

Abstract

As mechanical systems become more complicated and have diverse sub-modules, various sensor data are collected for the real-time health status monitoring of a system. However, because the collected sensor data are extremely large and contain irrelevant noise to the fault condition of the system, a technique of extracting important data fluctuations should be applied to detect the failure of the system. In general, unsupervised discretization techniques based on data distribution are used to extract fault patterns. However, the methods to extract significant features related to the state changes of a system are not simple. Therefore, we extract fault patterns by applying a supervised discretization method using not only the similarity between measurements but also the system state information. To verify the fault detection performance of the proposed method, acceleration sensor data were collected from a bearing-shaft system and analyzed using the proposed supervised discretized technique.