Deep Recurrent And Convolutional Networks For Accelerated Fault Tolerant Adaptive Flight Control Under Severe Failures

Design of fault tolerant systems is a popular subject in flight control system design. In particular, adaptive control approach has been successful in recovering aircraft in a wide variety of different actuator/sensor failure scenarios. However, if the aircraft goes under a severe actuator failure, control system might not be able to adapt fast enough to changes in the dynamics, which would result in performance degradation or even loss of the aircraft. Inspired by the recent success of deep learning applications, this work builds a hybrid recurrent/convolutional neural network model to estimate adaptation parameters for aircraft dynamics under actuator/engine faults. The model is trained offline from a database of different failure scenarios. In case of an actuator/engine failure, the model identifies adaptation parameters and feeds this information to the adaptive control system, which results in significantly faster convergence of the controller coefficients. Developed control system is implemented on a nonlinear 6-DOF F-16 aircraft, and the results show that the proposed architecture is especially beneficial in severe failure scenarios.