On-Board Neural Network Control of Kinetic Energy Projectiles for NEO Exploration

Locations and compositions of Near Earth Objects (NEO's) could be discovered by simple, inexpensive, kinetic energy projectiles launched by the thousands, provided they contain a minimum amount of propellent for course correction and a control system. The largest problem with these projectiles is incorporating enough advanced control capability on-board. The solution is the use of a compact, single chip, neural network control system. Our objective is to achieve more rapid and accurate tracking, coupled with better disturbance rejection, when compared to other controllers. The procedure was to classify a minimum list of cause and effect variables and train a recurrent neural network to mimic the dynamics of kinetic energy projectiles. We then incorporated the neural network into a model predictive control strategy, and computationally evaluated response times, tracking errors, and other performance indices. The results were excellent. Furthermore, because a neural network is inherently non-linear, it could be expanded to respond to non-linear effects not included in our simplified considerations. The neural network could also be trained in-flight. Our results were supported by on-line experimental results in a mining environment not directly related to kinetic energy projectiles. We concluded that a neural network architecture could provide optimal on-board control for propellant containing, kinetic energy projectiles, reducing the amount of control action taken, and minimizing the propellant required for course corrections.