Transient feature extraction based on phase space fusion by partial-least-square regression analysis of sensor array signals

Pattern classification based on transient signal analysis provides an effective method for identification of dynamical systems. The partial-least-square regression (PLSR) is most commonly used to generate parametric representation of phase space defined by measured signals and their time derivatives. The PLS component scores are interpreted as object features for pattern identification. In this paper, we consider sensor array transients, and propose PLSR based fusion of phase spaces of individual sensors into a single virtual phase space. Motivation for this approach comes from realizing that (i) multiplicity of array sensors encodes information about object diversity, and (ii) PLSR models object diversities in terms of small number of latent variables. The approach is validated through a case study of vapor identification by electronic nose based on surface-acoustic-wave (SAW) chemical sensor array. A comparison of results with and without fusion shows substantial improvement in vapor class separability after phase space fusion.