Propeller damage detection, classification and estimation in multirotor vehicles

TL;DR

The paper addresses propeller damage detection in multirotor UAVs by proposing a data-driven framework that uses only inertial measurements and commanded-torque data to detect, locate, and quantify damage. A two-stage pipeline combines a linear-kernel SVM for damage-type classification with neural networks that estimate damage magnitude and localize the damaged rotor, leveraging spectral-energy features across 22 bands plus torque statistics. Extensive indoor experiments, outdoor wind tests, and validation on the UAV-FD dataset demonstrate high accuracy for damage detection and localization, and robust magnitude estimation across symmetric, asymmetric, and longitudinal damage, with improvements over baselines. The findings suggest practical applicability for proactive maintenance and in-flight fault recovery, while showing generalization across vehicle topologies (quadrotor to hexacopter) and environmental conditions.

Abstract

This manuscript details an architecture and training methodology for a data-driven framework aimed at detecting, identifying, and quantifying damage in the propeller blades of multirotor Unmanned Aerial Vehicles. By substituting one propeller with a damaged counterpart-encompassing three distinct damage types of varying severity-real flight data was collected. This data was then used to train a composite model, comprising both classifiers and neural networks, capable of accurately identifying the type of failure, estimating damage severity, and pinpointing the affected rotor. The data employed for this analysis was exclusively sourced from inertial measurements and control command inputs, ensuring adaptability across diverse multirotor vehicle platforms.

Figures (18)

• Figure 1: Propellers with, A-longitudinal, B,C-asymmetrical, and D-symmetrical damages. The numbers on the right indicate the cuts performed on each of the 18 propellers used
• Figure 2: Multirotor with four rotor-propeller sets, with motor numbering, direction of thrust and torque of each actuator, and reference frames.
• Figure 3: Quadrotor platform used for flight data gathering, with one damaged propeller.
• Figure 4: Average power density over the frequency spectrum for the inertial sensors and the torque control signals for a flight with four healthy propellers (dashed). In transparency, the power density taken in several smaller sections of the same flight. The power density is almost the same, both if the vehicle is hovering, performing a maneuver, or when considering the full flight.
• Figure 5: Frequency spectrum for the inertial sensors and the torque control signals for a flight with a symmetrically cut propeller in rotor 1. Healthy flight as reference in dashed black line.