The main theme of this dissertation is networked dynamic nonlinear state estimation using EKF or UKF. The acquired sensor observations are time-stamped, and then transmitted over a digital communication network to a central estimation (and control) hub. The use of a communication medium introduces several challenges such as packet drops, irregularly-spaced observations, transmission delay, and bad data. Our continuous-discrete EKF and UKF can, by their very nature, handle arbitrarily-spaced sensor acquisition times. To counter some of the other challenges we augment the EKF/UKF with two modules: (i) a delay mitigation module, which uses additional storage to generate revised state estimates from delayed time-stamped measurements, and (ii) an innovations-based module that can detect the presence of a bad data measurement (and discard it). The dissertation results demonstrate the effectiveness of these modules in handling delays and corrupted sensor measurements.
To implement the continuous-discrete EKF/UKF algorithm in digital hardware, we discretize the continuous part of the system (i.e., the state equation), using Runge-Kutta explicit and implicit discretization schemes to reduce the cumulative error, as compared with the standard (Euler) discretization.
Finally, we use an ML-based iterative online tuning algorithm to reduce the effects of uncertain process and measurement noise covariance matrices.
- Professor Hanoch Lev-Ari (Advisor)
- Professor Bahram Shafai
- Professor Aleksander Stankovic