Multivariate Data Analysis for Motor Failure Detection and Isolation in A Multicopter UAV Using Real-Flight Attitude Signals

Reconfigurable aerial platforms such as multicopter unmanned aerial vehicles (UAVs) allow the design of fail-safe systems because of inherent redundancy in actuators and sensors to maintain stability with a reduction in flight performance. The methods based on univariate and multivariate time series analysis of just the attitude signals can pave the way for model-free systems that can be generalized across a class of UAVs like multicopters. In this paper, we present a critical analysis of real-flight attitude time-series signals and investigate them for data-driven motor fault and failure detection and isolation (FDI), specifically for multicopters configurations like quadcopters and hexacopters. We analyze flight data for different scenarios of outdoor flights, healthy and faulty, hovering and cruising, loss of efficiency, and single-rotor failure of every motor. We tested it for small to medium-sized multi-copters. The failure detection and classification are performed without relying on analytical system modeling or the knowledge of the controller. Thus, we perform three major assessments: vector autoregression (VAR) using residual variance, time-frequency analysis, and dimensionality analysis of the recorded variables, to support the classification framework. To the author's best knowledge, it is an early attempt at laying the foundation for engineering features from streaming attitude data, instead of simulations, that works on existing open-source autopilot hardware and is agnostic to the firmware as well. This foundation allows us to implement various FDI frameworks in real-time directly using the above variables on multicopters, which drastically increases the levels of safety and scalability of unmanned flights in drone applications.