Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


case1058208324.pdf (5.92 MB)

ETD Abstract Container

Abstract Header

An information based approach to anomaly detection in dynamic systems

Oh, Ki-Tae
http://rave.ohiolink.edu/etdc/view?acc_num=case1058208324

Abstract Details

1995, Doctor of Philosophy, Case Western Reserve University, Systems and Control Engineering.
An ultimate goal in dynamic system control is the development of a failure tolerant control system. Such systems require an anomaly detection mechanism and a reconfigurable controller. In this dissertation, a new anomaly detection method is developed for high speed detection. The problem is formulated mathematically in the observation space, and a hypothesized input-output model is assumed and associated with each anomaly class. The detection method developed is based on statistical information theory. The overall performance depends on the speed at which distinguishable features in the information are accumulating. Posterior distributions are calculated and used for the decision procedure. The speed of detection can be increased by effectively using the input signal to probe the system. This probing signal is synthesized in the time domain in feedback form. The main idea is to maximize the relative entropy between future output distributions of the two most plausible models. Modeling uncertainty is considered in the detection and in the synthesis of the probing signal. These detection mechanisms are extended to a Detection Network to handle a Nonlinear Non-Gaussian model. This is important for application purposes, because some anomaly classes may not be mathematically desc ribed using a Linear Gaussian model. In this Detection Network, a prediction module and a Probability Density Function (p.d.f.) module are used to obtain the necessary posterior distribution, and each module is built by using the Gaussian Basis Function Network (GBFN). In the p.d.f. module, the Maximum Likelihood estimation method incorporating Expectation Maximization algorithms is used for training, and various low complexity models in GBFN are considered. The Minimum Description Length Principle is applied to the p.d.f. model selection procedure to obtain a good compromise between simplicity and likelihood.
Kenneth Loparo (Advisor)
anomaly detection dynamic systems

Recommended Citations

Citations

  • Oh, K.-T. (1995). An information based approach to anomaly detection in dynamic systems [Doctoral dissertation, Case Western Reserve University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=case1058208324

    APA Style (7th edition)

  • Oh, Ki-Tae. An information based approach to anomaly detection in dynamic systems. 1995. Case Western Reserve University, Doctoral dissertation. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=case1058208324.

    MLA Style (8th edition)

  • Oh, Ki-Tae. "An information based approach to anomaly detection in dynamic systems." Doctoral dissertation, Case Western Reserve University, 1995. http://rave.ohiolink.edu/etdc/view?acc_num=case1058208324

    Chicago Manual of Style (17th edition)

case1058208324
541
© 1995, all rights reserved.
This open access ETD is published by Case Western Reserve University School of Graduate Studies and OhioLINK.