Multicopter Fault Detection and Identification via Data-Driven Statistical Learning Methods

This paper presents the introduction, investigation, and critical assessment of three data-driven methods for rotor failure detection and identification in a multicopter. These methods are based on aircraft attitude signals obtained from forward flight under turbulence and uncertainty. The knowledge-based method exploits the system rigid-body dynamics insight under the different rotor failures to construct a decision tree that detects and identifies the rotor failure simultaneously by how the roll, pitch, and yaw signals violate the statistical confidence limits immediately after failure. For the statistical time-series method, the development of stochastic time-series models and residual-based statistical hypothesis tests are discussed. Here, fault detection in the transient response is followed by identification after the signals reach a stationary state, after controller compensation, with the healthy and the different faulty models, respectively, in a two-step manner. The third method employs the healthy time-series model to extract a useful feature, which is the residual cross correlation, as an input to a neural network trained to achieve simultaneous rotor failure detection and identification. The time-series assisted neural network is capable of making decisions throughout the entire flight with an accuracy of 98.8%, with minimum computation time (less than 0.03 s) making it the best alternative for real-time monitoring.