Deep Learning-Based Monitoring of Flight Path Deviations for eVTOL Aircraft: A Gated Recurrent Unit Model Perspective

Abstract

—Urban Air Mobility (UAM) is transitioning from initial pilot-controlled deployments toward UAM Maturity Level 4, which envisions full autonomy in dense aerial environments. In such highly automated systems, even minor trajectory deviations can escalate into critical safety hazards. However, existing frameworks often lack the real-time precision and robustness required to detect these anomalies early, especially when faced with unexpected system disturbances or external interference. This study develops an artificial intelligence (AI)-based predictive model to identify deviations from planned flight paths and facilitate timely air tra ffi c control alerts. Using high-fidelity X-Plane simulations of the ALIA-250, an electric vertical take-o ff and landing (eVTOL) aircraft along Han River VFR routes, we collected autopilot data to train and compare Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) architectures. The GRU model demonstrated superior performance, achieving the lowest prediction error and 100% availability for a 3-second-ahead forecast. Furthermore, the model’s robustness was validated using synthetic flight data simulating various system failures and cyberattacks, confirming its potentialto provide early hazard warnings andsupporttransitions to manual controlwithin a simulated environment. These findingsdemonstrate the feasibility of AI-basedanomaly detectionas a proof-of-concept framework for trajectory monitoring in future UAM operations, with further validation under real-world conditions required before operational deployment.