This paper tackles the problem of optimal attitude control using a minimal number of attitude thrusters. Three possible control solutions to this problem are presented: 1) a logic-based controller that is simple to implement, 2) a projective control that aims to optimally replicate an ideal continuous control as closely as possible, and 3) an optimal neural predictive control (NPC) that minimizes the total impulse during a maneuver. The NPC is based on a recurrent neural network using a nonlinear autoregressive exogenous configuration for state propagation in a finite-time horizon optimization. Typically, for continuous systems, a back-propagation algorithm for the receding horizon optimization is used, but this is not applicable to systems with discrete inputs. In this paper the NPC is adapted to boolean input systems by employing a robust genetic algorithm to undertake the receding horizon optimization. An automatic selection of the parameters of the cost function is proposed, which improves the performance of the NPC and reduces the tuning to only one parameter. In addition, a multilayer perceptron (MLP) is trained offline with the optimal control data obtained, thus replacing the CPU-intensive cost function with a significantly less computationally expensive metamodel. The NPC performance is compared with the proposed logic-based and projective control algorithms in simulation of a 12U CubeSat and is shown to be the most efficient in terms of total impulse requirement at equal settling time and the least sensitive to the choice of parameters. The MLP control drastically reduces the online computational cost with performance approaching those of the NPC.
Neural-network-based optimal attitude control using four impulsive thrusters
Biggs J. D.;
2020-01-01
Abstract
This paper tackles the problem of optimal attitude control using a minimal number of attitude thrusters. Three possible control solutions to this problem are presented: 1) a logic-based controller that is simple to implement, 2) a projective control that aims to optimally replicate an ideal continuous control as closely as possible, and 3) an optimal neural predictive control (NPC) that minimizes the total impulse during a maneuver. The NPC is based on a recurrent neural network using a nonlinear autoregressive exogenous configuration for state propagation in a finite-time horizon optimization. Typically, for continuous systems, a back-propagation algorithm for the receding horizon optimization is used, but this is not applicable to systems with discrete inputs. In this paper the NPC is adapted to boolean input systems by employing a robust genetic algorithm to undertake the receding horizon optimization. An automatic selection of the parameters of the cost function is proposed, which improves the performance of the NPC and reduces the tuning to only one parameter. In addition, a multilayer perceptron (MLP) is trained offline with the optimal control data obtained, thus replacing the CPU-intensive cost function with a significantly less computationally expensive metamodel. The NPC performance is compared with the proposed logic-based and projective control algorithms in simulation of a 12U CubeSat and is shown to be the most efficient in terms of total impulse requirement at equal settling time and the least sensitive to the choice of parameters. The MLP control drastically reduces the online computational cost with performance approaching those of the NPC.File | Dimensione | Formato | |
---|---|---|---|
BIGGJ01-20.pdf
Accesso riservato
Descrizione: Paper
:
Publisher’s version
Dimensione
1.59 MB
Formato
Adobe PDF
|
1.59 MB | Adobe PDF | Visualizza/Apri |
BIGGJ_OA_01-20.pdf
Open Access dal 02/03/2020
Descrizione: Paper Open Access
:
Post-Print (DRAFT o Author’s Accepted Manuscript-AAM)
Dimensione
1.11 MB
Formato
Adobe PDF
|
1.11 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.