A Learning-Based Fault Detection Approach for Unmanned Aerial Systems Using Inertial Sensor Data

Abstract

This thesis presents a lightweight embedded fault detection framework for identifying propulsion system faults in Unmanned Aerial Vehicles (UAVs) using inertial sensor measurements. Mechanical degradation, imbalance, and physical damage affecting motors and propellers may reduce flight performance and compromise operational safety. Consequently, there is an increasing need for fault detection methods capable of operating in real time on resource-constrained embedded platforms.

The proposed approach utilizes three-axis acceleration and angular velocity measurements acquired from an onboard Inertial Measurement Unit. First, the raw sensor measurements undergo a lightweight preprocessing stage and are transformed into the frequency domain using the Fast Fourier Transform. Dominant frequencies, spectral amplitudes, band energies, and harmonic ratios are then extracted from the resulting spectrum to construct a compact feature vector.

To eliminate the need for labeled datasets, an online clustering algorithm is proposed for fault detection. The algorithm incrementally learns the characteristics of normal operating conditions and continuously evaluates newly acquired observations against the learned model. As a result, abnormal behavior can be identified without requiring an offline training stage while preserving the ability to adapt to changing operating conditions.

The proposed framework is experimentally evaluated under healthy, broken propeller, bent propeller, and direction-incompatible propeller scenarios. The results demonstrate that the method can successfully identify several abnormal operating conditions. Furthermore, power consumption, execution time, and memory analyses show that the proposed framework can operate in real time on resource-constrained embedded hardware.