Investigation of machine learning-based fault detection and severity estimation in unmanned aerial systems

Fault detection in unmanned aerial vehicles (UAVs) has become increasingly important due to their deployment in complex and safety-critical environments. Faults arising from components such as propellers, motors, and sensors can rapidly disrupt the tightly coupled control loops of multirotor UAVs, leading to instability, degraded performance, or complete loss of the vehicle. This underscores the need for effective fault detection, isolation, and severity estimation strategies that can recognize early signs of degradation. However, detecting such faults during real flights is challenging due to sensor noise, environmental disturbances, and the nonlinear behavior of multirotor aerodynamics.This thesis investigates a data-driven framework for rotor fault detection, localization, and severity estimation using convolutional neural networks (CNNs) trained on real flight data. The model operates by learning spatio–temporal patterns in the vehicle’s dynamic response, enabling it to infer both the affected rotor and the corresponding level of damage directly from flight dynamics. To generate data for training, controlled blade-tip breakages ranging from 0–20\% were systematically introduced to create realistic fault scenarios across all six rotors of a hexacopter. Flight experiments—consisting of steady hovers and full-trajectory missions—were conducted to produce multivariate time-series datasets capturing the vehicle’s kinematic and dynamic behavior under both nominal and faulty states.The model trained on these datasets effectively learned the underlying relationships between flight dynamics and rotor-level degradation, enabling simultaneous detection, isolation, and severity estimation. The results demonstrate strong predictive performance, with rotor-wise accuracies exceeding 99\% and an overall severity estimation accuracy of approximately 96\% when evaluated across all six rotors within a $\pm1\%$ damage tolerance range. Despite minor deviations—particularly at low absolute damage levels—the model consistently captured subtle damage signatures and generalized well across different fault configurations. In conclusion, this thesis provides an integrated approach that combines UAV design, development and fabrication, fault induction, real-world data acquisition, and machine learning–based model analysis. The resulting framework establishes a robust foundation for autonomous vehicle health monitoring and enhances the reliability of future UAV operations.