Intelligent Supervisory Control Architecture for Health Mon

Autonomous control of uninhabited air vehicles (UAVs) presents a number of challenges, including detecting failures across the entire flight regime/mission envelope, differentiating between behavioral changes due to failures and those due to uncertainties, real-time control law redesign, and real-time modification of UAV trajectories and sensor-allocation strategies. Barron Associates, Inc. proposes to address these challenges by developing supervisory control architectures that provide (1) robust failure detection across multiple regimes with relatively little a priori training and (2) on-line control redesign and mission replanning in the event of a failure. The proposed system uses a small number of performance models having parameters that vary as functions of flight regime. Multivariate model validation and statistical change detection algorithms are used to determine the validity of the model(s) given changing noise levels. On-line parameter identification is used to modify the parameters of some models and provide on-line reconfiguration capabilities. The best system model is used by (1) an inner-loop model-predictive that computes optimal control gains and (2) an outer-loop reinforcement learning supervisor that computes optimal sensor configurations and trajectory strategies for fulfilling mission objectives to the extent possible given the nature of unforeseen failure(s).