A Retrofit Architecture for Model-Based Adaptive Flight Control

Abstract

The paper presents an adaptive control framework that is integrated with the production control system of an in-service aircraft. The purpose is to maximize performance and safety for unexpected changes in the dynamics caused by flight control failures, damage, and adverse environmental conditions such as icing. Neural networks are developed from available high fidelity simulation and flight data to encode the dynamics of the nominal closed loop system. A structure learning modeling algorithm is used to address the model selection problem (terms and coefficients) of these neural network function approximators. A constrained parameter identification algorithm provides on-line model corrections that account for uncertainties or changes in the current aircraft dynamics, and the updated estimates are enlisted in a receding horizon optimal controller to provide increments to pilot commands. The increments from the adaptive control law serve to reduce tracking error given the current closed-loop characteristics of the aircraft. A key benefit of the approach is that the adaptation is only significant if the aircraft behavior differs appreciably from the intended closed-loop flying qualities. Furthermore, the control law reconfiguration is included through the control input paths and preserves the structural filters, mode logic, and custom performance and safety software of the original digital flight control system. The retrofit system was integrated with the production F/A-18 control augmentation system, and piloted simulations of inflight refueling, target tracking, and general maneuvering with unforeseen failures to primary aerodynamic control surfaces were performed by US Navy and Boeing pilots. The retrofit software is also implemented in the US Navy fleet support flight control computer, and, in general, real-time hardware in the loop test results support the findings of batch simulation and software-only piloted simulations. The retrofit reconfiguration architecture is summarized; the enabling neural network modeling and system identification methods are discussed, and an overview of the model-based control law is given. F/A-18 piloted simulations and hardware in the loop test results are provided to show the reconfiguration benefits of the method and to substantiate the claim of practical usefulness for fleet aircraft, respectively.