Identifying regions of interest in spectra for classification purposes

Distinguishing between separate classes of time-series data can often be simplified by using frequency-domain methods. Different states of a process can often result in different shapes and amplitudes in a spectral representation of the time series. The interpretation of the spectrum can be achieved in this case by identifying frequency regions which have a high discriminative power between the different classes, the so-called regions of interest (ROI). The discriminative power of two sequences is high if statistical or geometric parameters differ significantly between the classes. In this paper, a new approach for identifying these ROI, which makes use of image processing techniques is given. This new algorithm was developed in a research project with the aim of monitoring solenoid valves by analysing their mechanical vibration during switching on or off. Failure detection as developed in this project can be used for condition-based maintenance. However, it is anticipated that this new method can be generalised to many other similar types of classification problems. (C) 2002 Elsevier Science Ltd. All rights reserved.