AN ARTIFICIAL INTELLIGENCE APPROACH FOR FAULT DETECTION AND IDENTIFICATION ON CUBESAT PLATFORMS

Algorithms involving artificial intelligence (i.e. neural networks and fuzzy logics) have begun to spread in the space applications in recent years. This has been made possible thanks to the increase of the know-how on machine-learning techniques and the advance of the computational capabilities of on-board processing. Thanks to the advance of technology and miniaturization, the implementation of such techniques is now feasible even on small platforms, such as CubeSats. The paper presents an algorithm for magnetorquers fault detection and identification applied to a 3U CubeSat design. The implementation was developed in MathWorks Matlab & Simulink environment, following the model-based fault identification methodology. This approach involves artificial intelligence techniques, in particular fuzzy logic and multi-layer feed-forward offline-trained pattern recognition neural network. The simulated system is a CubeSat to fly a Low Earth Orbit mission: the attitude is controlled by three magnetorquers and a single reaction wheel, while the attitude determination is based on the measures from MEMS gyroscopes and magnetometers. Magnetorquers provide a reliable way to control attitude, and are one of the most used technologies for LEO CubeSats as, unlike other actuator options (e.g. thrusters), are relatively cheap, have low power consumption, and are lightweight. As any other component, they are subject of failures, whether temporary or definitive. Typical problems encountered by magnetorquers are: Float, Lock in Place, Hard-Over and Loss of Efficiency. Each of them alters in a peculiar way the behaviour of the actuator. The paper describes the simulator and algorithm architecture, with a particular focus on the design of fuzzy logics (connection and implication operators, rules and input/output qualificators) and the neural network architecture (number of layers, neurons per layer), threshold and activation functions, offline training algorithm and its data management. In conclusion, the paper presents the neural network and fuzzy states evolution and performance, in the different faulty configurations. Results show that the implementation of the fuzzy logics and neural networks helps to meet the specified performance requirements even in the event of some malfunctioning of a system actuator.