Spacecraft trajectory optimization using many embedded Lambert problems
Improvement of spacecraft trajectory optimization approaches, methods, and techniques is critical for better mission design. Preliminary low-fidelity analysis precedes high-fidelity analysis to efficiently explore the space of a problem. The work of this dissertation extends an embedded boundary value problem (EBVP) technique for preliminary design in the two-body problem. The EBVP technique is designed for direct, unconstrained optimization using many, short-arc, embedded Lambert problems that discretize the trajectory. The short arcs share terminal positions to implicitly enforce position continuity and the instantaneous velocity discontinuities in between segments are the control. These coasting arcs and impulsive maneuvers in between segments are defined collectively as a coast-impulse model, similar to the well-known Sims-Flanagan model.
Use of EBVPs is not new to spacecraft trajectory optimization, extensively used in primer vector theory, flyby-tour design, direct impulsive-maneuver optimization, and more. Lack of fast and accurate BVP solvers has prevented the use of the EBVP technique on problems with more than dozens of segments. For the two-body problem, a recently-developed Lambert solver, complete with the necessary partials, enables the extension of the EBVP technique to many hundreds to thousands of segments and hundreds of revolutions. The use of many short arcs guarantees existence and uniqueness for the Lambert problem of each segment. Furthermore, short arcs simultaneously approximate low thrust and eliminates the need to know the structure of a high-thrust impulsive-maneuver solution. A set of examples show the EBVP technique to be efficient, robust, and useful. In particular, an example using 256 revolutions, 6143 segments, and a constant flight time per segment, optimizes in 5.5 hours using a single processor.
After this initial demonstration, the EBVP technique is improved by a function which enables variable flight time per segment. Guided by the well-known Sundman transformation, these piecewise Sundman transformation (PST) functions divide the total flight time of the trajectory into spatially-even arcs, importantly not modifying the dynamics. Flight-time functions and their dynamical regularization counterpart are shown to share similar behavior for Keplerian orbit propagation. The PST functions are also shown to extend the EBVP technique to a large design space, where a runtime-feasible transfer with 512 revs and 12287 segments is presented that significantly changes semimajor axis, eccentricity, and inclination. Moreover, another example is presented that transfers through the numerically challenging parabolic boundary, i.e. a transfer from a circular to hyperbolic orbit. Both these examples use an exponent of 3/2 for the PST to enforce the spatially-even arcs or equal steps in eccentric anomaly.
Lastly, an optimal control problem is formulated to solve a class of many-revolution trajectories relevant to the EBVP technique. For transfers that minimize thrust-acceleration-squared, primer vector theory enables the mapping of direct, many-impulsive-maneuver trajectories to the indirect, continuous-thrust-acceleration equivalent. The mapping algorithm is independent of how the direct solution is obtained and the mapping computations only require a solver for a BVP and its partial derivatives. For the two-body problem, a Lambert solver is used. The mapping is simple because the impulsive maneuvers and co-states share the same linear space around an optimal trajectory. For numerical results, the direct coast-impulse solutions are demonstrated to converge to the indirect continuous solutions as the number of impulses and segments increase. The two-body design space is explored with a set of three many-revolution, many-segment examples changing semimajor axis, eccentricity, and inclination. The first two examples change either a small amount of semimajor axis or eccentricity, and the third example is a transfer to geosynchronous orbit. Using a single processor, the optimization runtime is seconds to minutes for revolution counts of 10 to 100, while on the order of one hour for examples with up to 500 revolutions. Any of these thrust-acceleration-squared solutions are good candidates to start a homotopy to a higher-fidelity minimization problem with practical constraints.