ITLApplied  Computational Mathematics Division
ACMD Seminar Series
Attractive Image NIST

Anomaly Detection and Failure Mitigation in Complex Dynamical Systems

Asok Ray
Pennsylvania State University, Department of Mechanical and Nuclear Engineering

Friday, November 19, 2004 11:00-12:00,
NIST North (820), Room 145
Friday, November 19, 2004 09:00-10:00,
Room 4550

Abstract: Engineering theories of control, communication, and computation have matured in recent decades facilitating creation of systems of bewildering complexity, which are almost comparable to biological systems. The complexity is often hidden and cryptic in the laboratory environment as well as during nominal operations of large-scale dynamical systems; however, it may become acutely conspicuous when contributing to rare cascading faults. From these perspectives, anomaly is defined as deviation from the nominal behavior and is associated with parametric or non-parametric changes that may gradually evolve. Early detection of anomalies in complex dynamical systems, such as future-generation airborne vehicles, is essential not only for prevention of catastrophic failures and mission disruption, but also for robust performance, life extension, and self healing. The seminar will narrate some of the research experience on Anomaly Detection and Failure Mitigation in Complex Dynamical Systems under a current Multidisciplinary University Research Initiative (MURI) grant from the Army Research Office (ARO). The first part of the seminar will focus on early detection of potentially malignant anomalies in real time, based on the available time series data of macroscopic observables. The second part of the seminar will use this information for failure mitigation and accommodation via discrete-event supervisory control. The proposed concept of anomaly detection is built upon the theories of Statistical Mechanics, Symbolic Time Series Analysis, Pattern Recognition, and Computational Mechanics. The key idea is identification of anomaly patterns from symbolic sequences, derived from time series data, through finite-state machines having the structure of the generalized Ising (Potts) model; this is accomplished by taking advantage of non-linear and non-stationary features of the dynamical system. The underlying principle of anomaly detection is analogous to the canonical ensemble theory of microstates under thermodynamic equilibrium for quasi-stationary behavior as well under phase transitions. The anomaly detection concept will be first illustrated using an example of a second-order forced Duffing equation where the dissipation parameter is slowly varying and small parametric changes are captured much before the onset of chaotic motion (i.e., period doubling) at a bifurcation point. Then, the efficacy of anomaly detection will be demonstrated on experimental data sets of fatigue crack damage in polycrystalline alloys. The concept and formulation of a signed real measure of regular languages will then be presented for synthesis of discrete-event robust optimal supervisors that use the anomaly information. The measure is constructed based upon the principles of automata theory and real analysis for quantitative evaluation and comparison of the controlled behavior of discrete-event dynamical systems. The language measure creates a total ordering on the performance that provides a precise quantitative comparison of the plant behavior under different supervisors. Total variation of the language measure serves as a metric for the vector space of sublanguages of the regular language.

Speaker Bio: Asok Ray earned the PhD degree in Mechanical Engineering from Northeastern University in 1976, as well as graduate degrees in Electrical Engineering, Mathematics, and Computer Science. Dr. Ray joined Pennsylvania State University in 1985 and is currently a Distinguished Professor of Mechanical Engineering, a Graduate Faculty of Electrical Engineering, and a Graduate Faculty in the Inter-College Program in Materials. Prior to joining Pennsylvania State University, Dr. Ray held research and academic positions at Massachusetts Institute of Technology and Carnegie-Mellon University as well as management and research positions at GTE Strategic Systems Division, Charles Stark Draper Laboratory, and MITRE Corporation.

Presentation Slides: PDF

Contact: S. Langer

Note: Visitors from outside NIST must contact Robin Bickel; (301) 975-3668; at least 24 hours in advance.

Privacy Policy | Disclaimer | FOIA
NIST is an agency of the U.S. Commerce Department.
Last updated: 2011-01-12.