Autolanding control system design with deep learning based fault estimation

Developing a control system that can recover the aircraft under actuator failures and severe disturbances is a popular problem in flight control system design. Majority of existing controllers cannot adapt to changes in aircraft dynamics that occur due to severe actuator failures. We propose a novel, data-driven fault estimation method for estimating actuator faults from aircraft state trajectories, which utilizes a deep neural network trained offline on gathered simulation data of fault injected aircraft. Proposed novel deep fault estimation model coupled with an existing nonlinear dynamic inversion based autolanding controller, reacts immediately to a wide range of actuator failures and is able to land the aircraft under many different combinations of actuator failures and severe wind conditions. Performance of the proposed approach is compared to existing state-of-the-art fault tolerant controllers through fault tolerance maps. It is observed that the developed approach is superior both in terms of fault tolerance map coverage and smoothness of the computed controller signals.