Using learning techniques to accommodate unanticipated faults

Abstract

Procedures exist to rapidly accommodate certain types of faults, based on a priori specification of the postfault dynamics. A methodology is presented for accommodating the remaining unanticipated faults. For these two approaches, a tradeoff exists between the time to attain a solution to the reconfiguration problem and the generality of the approach. Unanticipated faults are represented as unmodeled forces and torques. Models of these forces and torques are developed online using a hybrid estimation/learning approach. The hybrid system is designed for fast estimation during the initial transient when a fault occurs, with continually improving performance as postfault information is accumulated by the learning system. Fault accommodation is achieved by a feedforward/feedback control architecture that employs an actuator distribution system to convert desired forces into individual actuator commands. This approach is demonstrated on a simulated autonomous vehicle, where the addition of a hybrid estimation/learning capability is shown to increase performance greatly over time. >