Deep recurrent and convolutional networks for robust fault tolerant autonomous landing control system design under severe conditions

Developing a control system that is tolerant to actuator and sensor faults is one of the major problems in flight control system design. In many failure scenarios, it is too dangerous to continue the mission and hence aircraft is ordered to perform an emergency landing. Thus, it is critical that an autonomous landing system should be capable of landing the aircraft under severe actuator and sensor failures and external disturbances such as wind. In this paper, based on previous work, we present a robust nonlinear dynamic inversion based landing control system that can accommodate actuator failures and wind disturbances. We further improve the performance of the system by using a deep recurrent network to estimate the air-data parameters such as angle of attack, when the pitot tube measurements are unavailable/noisy. Simulation results show that the developed system is able to land the aircraft safely for a wide variety of failure and wind disturbance scenarios. In particular, use of a deep neural network makes a considerable difference in severe wind conditions.