Studying the Efficiency of Hybrid Algorithms for Fault Detection and Isolation in Navigation Sensors

Detection and isolation of faults in measurements of navigation sensors is considered. This problem is relevant in the context of autonomous navigation, since in this case, fault-tolerance requirements for the equipment generating navigation solutions are much higher. The paper proposes using both classical techniques based on state observers and machine learning methods. The efficiency of the performance of models based on random forest and recurrent neural network, as well as a hybrid version involving additional data from sliding mode observers are compared. The initial data for training and testing is a generated dataset based on a nonlinear model of a large-capacity vessel motion in the horizontal plane, which takes into account external disturbances, as well as errors of navigation sensors. The class of faults considered is limited to additive faults in the measurement equation.