Fault detection and diagnosis system considering aerodynamic effects for unmanned aerial vehicles.

Abstract

Attitude control in quadrotors usually does not consider the possibility of actuator lossof-effectiveness, becoming then inefficient at certain operating conditions. Fault-Detection and Diagnosis (FDD) algorithms available in the literature do not take into consideration the presence of aerodynamic effects such as blade flapping and induced drag, which hinders flight performance. The attenuation these effects cause on actuator speed could even be mistaken for a loss-of-effectiveness fault, so an FDD system capable of identifying them could be opportune. Typical quadrotor control systems as a whole tend to neglect the presence of such aerodynamic effects. As a primary effort, this thesis proposes that these effects can be compensated in the control allocation step, using a function-fitting neural network as a tool to replace the classic allocation matrix, without using the aerodynamic inflow equations directly. The network training is performed offline, saving on computational power. The target system is a PARROT Mambo drone. This specific quadrotor is particularly susceptible to the aerodynamic effects of interest to this thesis, given its small size. In this sense, when comparing the commands sent by the controller that must be achieved by the actuators, to the mechanical torques generated by them, the usage of the proposed neural network control allocation (NNCA) makes it possible to achieve a close match, meaning the actuators execute the appropriate control effort demanded by the controller, while the classic allocation matrix cannot perform the same way. Furthermore, the closed-loop performance was also improved with the use of the new control allocation, as well as the quality of the thrust and torque signals, in which a much less noisy behavior was perceived. In order to be coherent with the inclusion of the aerodynamic effects modeling into the system through the control allocation, these equations are also included in the FDD formulation. The new fault detection scheme is based on a Reduced Order EKF Tolerant to Aerodynamic Effects (ROEKF-TAE) capable of detecting loss-of-effectiveness (LOE) faults in more than one actuator at the same time. With this goal, a more complete model of the quadrotor system including the pertinent aerodynamic effects equations is introduced in the filter’s development. Moreover, the previously developed NNCA was applied in all iv the new FDD scheme simulations and experiments. The experimental flights for the new FDD approach were performed using the same PARROT Mambo micro drone. They were compared to those of a traditional EKF, as well as to those of a state-of-the-art adaptive Kalman Filter available in the literature. The ROEKF-TAE is able to better distinguish between the effects of the actual fault and those of the blade-flapping disturbance. Also, it has better accuracy in identifying which actuators are truly defective and which are not.