ORCID Number
0009-0009-5740-9820
Date of Award
Summer 2026
Access Type
Dissertation - ERAU Login Required
Degree Name
Doctor of Philosophy in Aerospace Engineering
Department
Aerospace Engineering
Committee Chair
Troy Henderson
Committee Chair Email
First Committee Member
Morad Nazari
First Committee Member Email
Second Committee Member
Hao Peng
Second Committee Member Email
Third Committee Member
Yongxin Liu
Third Committee Member Email
College Dean
James W. Gregory
Abstract
Heading estimation is a foundational problem in planetary descent and other GPS-denied navigation settings, since terrain-relative navigation, hazard avoidance, and guidance all depend on it as a reference frame. Classical Kalman filter variants remain the standard tool for this problem, but their accuracy depends on assumptions about vehicle dynamics and noise statistics that can break down during high-dynamic flight. This dissertation develops and evaluates a heading and attitude estimation framework that compares model-based Kalman filtering against a data-driven Long Short-Term Memory (LSTM) network, trained independently and evaluated on equal footing, across four physically distinct platforms: a suborbital reusable launch vehicle, a solid-propellant sounding rocket, a legged ground robot, and a free-flying spacecraft.
Heading is propagated from gyroscope-measured angular rate with a co-estimated bias state, and corrected using two complementary IMU-derived observables, one from integrated velocity and one from gravity-compensated acceleration. Linear, Extended, and Unscented Kalman Filter formulations are compared against an LSTM trained end to end on raw six-channel IMU data, with no filtered quantity shared between the two approaches. A hodograph-based geometric validation method is introduced to check the physical consistency of estimated heading trajectories without requiring ground truth.
Across all four case studies, the LSTM consistently demonstrates strong in-sample accuracy, but its performance on held-out data is closely tied to how well the training data covers the conditions being evaluated. On the suborbital flight, held-out LSTM error is roughly two hundred times larger than its in-sample error, while a hodograph check confirms the network still produces a physically consistent heading trajectory. For the ground robot, the LSTM achieves more than an eighty percent improvement over the classical filter on well-represented maneuvers, while its performance degrades for maneuvers with limited representation in the training data.
For the free-flying spacecraft, the independently formulated MEKF and QUKF produce nearly equivalent attitude-error performance across the evaluated trajectories. Their close agreement suggests that, for this dataset, performance is influenced more by the information and trajectory coverage available in the measurements than by the choice between the two filter formulations. The LSTM achieves substantially lower attitude error under matched training conditions but exhibits a similar loss of performance on held-out trajectories, indicating sensitivity to the limited duration and diversity of the available training data.
Taken together, these results indicate that the generalization gap observed in this dissertation is a consequence of limited and narrow training data rather than a limitation of the underlying approach. The classical Kalman filter family remains the more reliable choice for operational heading estimation under the conditions examined here, but the evidence collected across four platforms supports continued development of data-driven methods as a complement to, rather than a replacement for, model-based filtering in future AI-assisted navigation systems.
Scholarly Commons Citation
Patel, Khushboo, "Toward AI-Assisted Heading Estimation: A Comparative Study of Kalman Filtering and Recurrent Neural Networks Using IMU Data" (2026). Doctoral Dissertations and Master's Theses. 1024.
https://commons.erau.edu/edt/1024
