Statistical Residual-Based Time Series Methods for Multicopter Fault Detection and Identification

This paper presents the introduction, investigation, and critical assessment of three data-driven statistical residual-based time series methods for rotor fault detection and identification (FDI) in multicopters. A concise overview of statistical residual-based FDI methods is provided based on scalar (univariate) and vector (multivariate) stochastic time series models. The FDI methods employed in this study are based on identified response-only parametric scalar (univariate) and vector (multivariate) autoregressive (AR) representations of multicopter attitudes (time series), as the external excitation is non-observable, and their corresponding model residuals obtained under the considered healthy and faulty multicopter states. The comparative assessment of the effectiveness of three residual-based statistical FDI methods is presented in the face of external disturbances, namely three different levels of turbulence, and for different rotor fault scenarios. To the authors' best of knowledge, this is the first time that residual-based statistical time series methods are investigated and evaluated with respect to multicopter FDI. The unique characteristics of the presented methods are: (i) the data-driven stochastic identification of the multicopter dynamics under healthy and faulty rotor conditions is based only on the use of multicopter attitude signals, i.e. roll, pitch, and yaw, without the need to use additional aircraft states; (ii) there is no analytical modeling involved and the subsequent development of a stochastic multicopter FDI framework does not require knowledge of the controller effort to detect and classify rotor faults. The fault detection methods based on Vector AutoRegressive (VAR) models exhibit improved performance compared to their scalar counterparts, as indicated by lower false alarms and missed fault rates. In the case of rotor fault identification (classification), the methods that are based on scalar AR models exhibit reduced rotor fault classification accuracy, while the VAR-based methods outperform their scalar counterparts and can achieve a fault identification accuracy of up to 99.6%.