An explainable stacking ensemble for UAV propeller fault diagnosis using electromechanical flight telemetry

Résumé fourni par la source

: Unmanned aerial vehicle (UAV) fault diagnosis is an important problem at the intersection of machine learning (ML), signal processing, and intelligent control systems. This study presents an explainable artificial intelligence framework for multiclass fault diagnosis in UAVs using controlled flight telemetry and stacked ML. A curated subset of the publicly available DronePropA dataset was used, incorporating inertial measurement unit (IMU) and electronic speed controller (ESC) signals acquired under healthy and defective propeller conditions. To characterize fault-related variations in electromechanical dynamic response, a structured feature engineering pipeline was developed to extract time-domain, frequency-domain, statistical, energy-based, correlation, and actuation descriptors from 1 kHz telemetry signals. Three baseline classifiers, namely k-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Stochastic Gradient Descent (SGD), were integrated within a stacking ensemble to improve classification robustness across four operational states: Healthy, Edge-Cut, Crack, and Surface-Cut. Three evaluation scenarios were conducted. Under the controlled baseline using Drone 1, trajectory t1, and the high-speed condition, the stacking ensemble achieved 94.8% accuracy, 94.2% precision, 93.7% recall, and a 93.9% F1-score. In the external healthy-state evaluation using unseen Drone 2 telemetry, the ensemble achieved 98.3% accuracy, 100.0% precision, 96.2% recall, and a 98.1% F1-score, demonstrating strong recognition of healthy operation across platforms. To enhance model transparency, SHAP analysis was applied to quantify feature contributions, revealing that motor command variance, signal energy, and IMU cross-correlation were the three most influential predictors in the global feature-importance ranking. The results demonstrate that explainable AI can provide accurate fault diagnosis across trajectory-variable and cross-drone conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.