Stall Prediction in Quadcopters With SHAP-Based Explainability and a Novel Flight Dynamics Dataset

Abstract

Uncrewed Aerial Vehicles (UAVs), particularly quadcopters, are increasingly deployed in high-stakes operations including defense, tactical surveillance, disaster response, and GPS-denied missions. However, aerodynamic stall—caused by conditions like vortex ring state or rotor stall—remains a critical failure mode that can lead to sudden lift loss and catastrophic crashes. This study presents a hybrid stall prediction framework that unifies physics-informed synthetic data generation, interpretable machine learning, and embedded rule-based logic for real-time, onboard risk assessment. A custom simulator was developed to generate 25,921 stall-prone flight states by modeling key parameters such as throttle, vertical speed, blade angle of attack (AoA), airspeed, disc loading, and thrust-to-weight ratio. Using this dataset, an XGBoost classifier was trained for binary classification with two output classes: stall and non-stall, and optimized via threshold tuning and SMOTE-based class balancing. Using this dataset, an XGBoost classifier achieved 0.97 precision, 0.98 recall, 0.98 F1-score, and 0.96 accuracy for stall prediction after integration with domain-specific rules. SHapley Additive exPlanations (SHAP) values were applied to provide transparent, instance-level justifications, revealing that throttle, VRS, and disc loading were consistently the top contributors to stall predictions. In parallel, domain-specific aerodynamic rules—such as critical AoA and high disc loading thresholds—were deployed to catch edge cases such as near-threshold predictions, ambiguous vortex ring state regions, or unseen aerodynamic configurations. This modular system is designed for compatibility with onboard processors like Raspberry Pi, enabling real-time deployment even in resource-constrained environments. It supports frugal innovation without compromising safety, making it suitable for stealth drones, ISR swarms, payload-carrying UAVs, and companion-computer-based quadcopters. By combining data-driven predictions with physics-aware rule triggers, the system reduces false negatives and false positives, improving reliability in mission-critical conditions. Ultimately, this work pushes drone autonomy toward proactive, explainable, and hardware-efficient safety mechanisms—where UAVs not only predict aerodynamic stall but also justify it in real time.