Fault Detection and Isolation for a Standard Quadrotor Using a Deep Neural Network Trained on a Momentum-based Estimator

Rapid identification of motor failures holds significant importance for ensuring the safety of multi-rotor unmanned aerial vehicles. This study introduces a method for detecting and isolating motor faults in standard quadrotors, utilizing an external wrench estimator and a recurrent neural network with long short-term memory nodes. The proposed approach treats partial or complete motor failure as an external disturbance affecting the quadrotor. Consequently, the external wrench estimator trains the network to quickly discern whether the estimated wrench results from a motor fault, identifying the specific motor involved, or if it stems from unmodeled dynamics or external factors such as wind or contacts. The training and testing of this method were conducted in a simulation environment equipped with a physics engine.