Time-series Assisted Machine Learning Framework for Probabilistic Rotor Fault Diagnosis on Multicopters under Varying Operating Conditions

A novel machine-learning probabilistic framework for online rotor fault detection, identification, and quantification in multicopters via strain signals is introduced. The framework performs robustly under varying flight states, i.e. forward velocity and gross weight configurations, as well as effectively accounts for the effects of gusts. It employs in-flight time-series strain data obtained from a 2-feet diameter hexacopter flying under external disturbances and uncertainty. The proposed scheme relies on out-of-plane strain measurements at each of the multicopter booms to diagnose, i.e. detect, identify and quantify, rotor faults while distinguishing them from the aircraft response to random gusts. A simple perceptron is shown to be both effective and robust for performing simultaneous online rotor fault detection and identification. Next, linear regression models are used to predict the rotor degradation value with 95% confidence intervals using strain data at the boom on which the faulty rotor is mounted. Indicative results for test operating conditions (not used in the training phase) demonstrate the generalization capability of the method. The proposed framework can accurately detect, identify and quantify minor rotor faults of 10\% degradation while distinguishing them from aggressive gusts of up to 10 m/s magnitude. The maximum time of fault detection is less than 0.3 s while achieving classification and quantification accuracy over 99%.