Industrial Applications of Extended Output-Only Blind Source Separation Techniques

Abstract

In the field of structural health monitoring or machine condition monitoring, most vibration based methods reported in the literature require to measure responses at several locations on the structure. In machine condition monitoring, the number of available vibration sensors is often small and it is not unusual that only one single sensor is used to monitor a machine. This paper presents industrial applications of two possible extensions of output-only Blind Source Separation (BSS) techniques, namely Principal Component Analysis (PCA) and Second Order Blind Identification (SOBI). Through the use of block Hankel matrices, these methods may be used when a reduced set of sensors or even one single sensor is available. The objective is to address the problem of fault detection in mechanical systems using subspace-based methods. The detection is achieved by comparing the subspace features between the reference and a current state using the concept of angular coherence between subspaces.