COMPLEX SYSTEMS
Current Research - Forewarning of Structural Failure from Time Series Analysis
One of the most important problems encountered in the study of complex, distributed systems is the appropriate characterization, detection, and prognosis of significant changes in the underlying dynamics. This problem is especially difficult when no model is available. Global features of complex dynamics may be extracted from time series alone and quantified by certain nonlinear descriptors. However, these descriptors are not sufficiently sensitive to distinguish between different chaotic regimes, especially when data are limited and/or noisy.
To address this problem, we introduce new metrics. Time-serial data is converted to a distribution function (DF) which follows with higher fidelity the changes in the dynamics. Different situations are compared via new "distances" between the DFs, which are correspondingly more discriminating. Large dissimilarity measures signify that the system has departed from the base case and forewarns of an impending unusual event, in particular a structural failure. The superiority of our methodology in forewarning such events is illustrated here for an electric motor seeded fault. While traditional nonlinear measures vary erratically (A), the dissimilarity measures show unambiguous trends as the motor undergoes the slow transition from normal regime to failure.
Research by V. Protopopescu and L. M. Hively: Phys. Lett A 258 103 (1999), CHAOS 10 864 (2000), JOCN 18 223 (2001), various Conference Proceedings.
|
||||||
[