ABSTRACT Unmanned Aerial Vehicles (UAVs) rely on tightly coupled onboard sensors for stable and safe flight; however, early-stage fault detection remains challenging due to noisy telemetry, nonlinear dynamics, and limited labelled fault data. This study presents a simulation-driven, explainable machine-learning framework for multi-sensor UAV fault diagnosis within digital-twin environments. A physics-informed telemetry generator produces synchronized accelerometer, gyroscope, motor RPM, battery voltage, and GPS signals with systematic injection of representative fault modes, including motor degradation, battery sag, IMU drift, and GPS anomalies. The multivariate time series is segmented using overlapping windows and transformed into a 60-dimensional statistical feature space designed for interpretability and computational efficiency. A Random Forest classifier achieves 99.33% test accuracy and an AUC of 0.997 under controlled simulation conditions. SHAP-based explainability quantifies feature-level contributions, highlighting the dominant influence of gyroscope drift and battery-voltage descriptors consistent with known degradation mechanisms. The framework provides a lightweight, reproducible, and interpretable diagnostic pipeline for simulation-based verification and early-stage UAV health-monitoring development.
Aswin Karkadakattil (Mon,) studied this question.