HR: 12:10h
AN: NG22A-08    [Abstracts]
TI: Phase-space dissimilarity measures for industrial and biomedical applications
AU: * Protopopescu, V A
EM: vvp@ornl.gov
AF: Oak Ridge National Laboratory, 1, Bethel valley Rd., Oak Ridge, TN 37831 United States
AU: Hively, L M
EM: hiv@ornl.gov
AF: Oak Ridge National Laboratory, 1, Bethel valley Rd., Oak Ridge, TN 37831 United States
AB: One of the most important problems in time-series analysis is the suitable characterization of the dynamics for timely, accurate, and robust condition assessment of the underlying system. Machine and physiological processes display complex, non-stationary behaviors that are affected by noise and may range from (quasi-)periodic to completely irregular (chaotic) regimes. Nevertheless, extensive experimental evidence indicates that even when the systems behave very irregularly (e.g., severe tool chatter or cardiac fibrillation), one may assume that - for all practical purposes - the dynamics are confined to low dimensional manifolds. As a result, the behavior of these systems can be described via traditional nonlinear measures (TNM), such as Lyapunov exponents, Kolmogorov entropy, and correlation dimension. While these measures are adequate for discriminating between clear-cut regular and chaotic dynamics, they are not sufficiently sensitive to distinguish between slightly different irregular (chaotic) regimes, especially when data are noisy and/or limited. Both machine and physiological dynamics usually fall into this latter category, creating a massive stumbling block to prognostication of abnormal regimes. We present here a recently developed approach that captures more efficiently changes in the underlying dynamics. We start with process-indicative, time-serial data that are checked for quality and discarded if inadequate. Acceptable data are filtered to remove confounding artifacts (e.g., sinusoidal variation in three-phase electrical signals or eye-blinks and muscular activity in EEG). The artifact-filtered data are then used to recover the essential features of the underlying dynamics via standard time-delay, phase-space reconstruction. One of the main results of this reconstruction is a discrete approximation of the distribution function (DF) on the attractor. Unaltered dynamics yield an unchanging geometry of the attractor and the visitation frequencies of its various points, corresponding to the baseline DF. Condition change is established by comparing the base line DFs to subsequent test case DFs via new, phase space dissimilarity measures (PSDM), namely the distance and - square statistics between two DFs. A clear trend in the dissimilarity measures over time indicates substantial departure from the baseline dynamics, thus signaling condition change. The severity of this departure can be interpreted as a "normal" fluctuation, abnormal behavior, impending failure, or complete breakdown. We illustrate the new approach on an assortment of machinery and biomedical examples. The machine data were collected during laboratory tests on industrial equipment, for diverse failure modes, via seeded faults and accelerated failures. The biomedical applications involve detection of physiological changes, such as epileptic seizures from EEG; ventricular fibrillation, fainting, and sepsis onset from ECG; and breathing difficulty from chest sounds. The PSDM show a consistent discrimination of normal-to-abnormal transitions, allowing earlier, more accurate, and more robust detection of the dynamical change for all of these applications in comparison to TNM.
DE: 3238 Prediction (3245, 4263)
DE: 3270 Time series analysis (1872, 4277, 4475)
DE: 4430 Complex systems
SC: Nonlinear Geophysics [NG]
MN: Fall Meeting 2005