Control of Rendez-vous and Docking phases of two CubeSats, by means of artificial intelligence algorithms.

On-orbit proximity operations require advanced guidance, navigation and control systems to obtain high pointing accuracy and very precise manoeuvring. The design of an attitude and orbit control is of primary importance in such delicate operations, since a spacecraft is subjected to many disturbance forces and torques, due to many sources, that lead to orbit variation (e.g. decay) and undesired angular accelerations. For missions with proximity operation, innovative Attitude Determination and Control Subsystem (ADCS) and a Guidance, Navigation and Control Subsystem (G, N &S) are of paramount importance. Orbit and attitude control is carried out thanks to the proper interactions of several elements, such as sensors, controller and actuators. The sensors determine the position and orientation of the body in space, but also its translational speed and rotational velocity around its centre of mass, with respect to a local and inertial reference systems. The control system is governed by sophisticated algorithms that aim at optimize the number and quality of manoeuvres and fuel, determining the commands to actuators, in order to correct the spacecraft trajectory and orientation and match the desired values. The complexity of this system increases in the small satellite field, where the technology is not completely mature and these kind of missions are never been performed. The aim of this thesis is the definition of a control strategy for proximity manoeuvres between two 6U CubeSats. In particular, it was analysed the control of the last phases of the rendez-vous, from the last hold point up to the mating, the rendez-vous and docking operations, through the definition of the mathematical model of the problem (i.e. relative translational and rotational dynamics and kinematics, environmental effects on the spacecraft movements), its implementation and simulation in Matlab/Simulink, and the assessment of the results, considering different simulation conditions (e.g. final accuracy, execution time, initial position and orientation). The choice of the control strategy to adopt in this work fell on machine learning algorithms; more precisely, a direct artificial neural network control was developed, in order to explore the capability . Artificial neural networks are a framework of many different machine learning algorithms and they are a mathematical and simplified representation of a biological neural system. Their main characteristic is the learning capability from a training input/output set or a mathematical function: the neural network, once learnt its task, can reproduce the desired output, even if the inputs change. The neural network was trained with an input/output set of values acquired by the simulation of the same model in which the control was exerted through a Proportional-Integral-Derivative controller. The ability of neural network algorithms to adapt to different and sometimes unpredictable conditions could be very useful in spacecraft control problems, and has been studied in this work. In particular, a basic comparison between two controller, one based on neural networks and one Proportional Integrative and Derivative controller, was made demonstrating the more adaptability and efficiency of the neural network method over PID control, when external conditions change.