Input and State Estimation for Discrete-Time Linear Systems with Application to Target Tracking and Fault Detection

Abstract

This dissertation first presents a deterministic treatment of discrete-time input reconstruction and state estimation without assuming the existence of a full-rank Markov parameter. Algorithms based on the generalized inverse of a block-Toeplitz matrix are given for 1) input reconstruction in the case where the initial state is known; 2) state estimation in the case where the initial state is unknown, the system has no invariant zeros, and the input is unknown; and 3) input reconstruction and state estimation in the case where the initial state is unknown and the system has no invariant zeros. In all cases, the unknown input is an arbitrary deterministic or stochastic signal. In addition, the reconstruction/estimation algorithm is deadbeat, which means that, in the absence of sensor noise, exact input reconstruction and state estimation are achieved in a finite number of steps.

Next, asymptotic input and state estimation for systems with invariant zeros is considered. Although this problem has been widely studied, existing techniques are confined to the case where the system is minimum phase. This dissertation presents retrospective cost input estimation (RCIE), which is based on retrospective cost optimization. It is shown that RCIE automatically develops an internal model of the unknown input. This internal model provides an asymptotic estimate of the unknown input regardless of the location of the zeros of the plant, including the case of nonminimum-phase dynamics.

The input and state estimation method developed in this dissertation provides a novel approach to a longstanding problem in target tracking, namely, estimation of the inertial acceleration of a body using only position measurements. It turns out that, for this problem, the discretized kinematics have invariant zeros on the unit circle, and thus the dynamics is nonminimum-phase. Using optical position data for a UAV, RCIE estimates the inertial acceleration, which is modeled as an unknown input. The acceleration estimates are compared to IMU data from onboard sensors.

Finally, based on exact kinematic models for input and state estimation, this dissertation presents a method for detecting sensor faults. A numerical investigation using the NASA Generic Transport Model shows that the method can detect stuck, bias, drift, and deadzone sensor faults. Furthermore, a laboratory experiment shows that RCIE can estimate the inertial acceleration (3-axis accelerometer measurements) and angular velocity (3-axis rate-gyro measurements) of a quadrotor using vision data; comparing these estimates to the actual accelerometer and rate-gyro measurements provide the means for assessing the health of the accelerometer and rate gyro.