Proactive Early Warning of Vortex Ring State in Coaxial UAVs: A Physics-Informed Multimodal ViT-LSTM Approach

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Vortex Ring State (VRS) presents a critical aerodynamic risk to coaxial dual-rotor UAVs.
• Existing detection methods are reactive or suffer from data limitations like leakage and imbalance.
• Lack of physical interpretability hinders trust in current data-driven diagnostic systems.

Purpose of the Study:

• To develop a proactive early warning system for VRS in UAVs.
• To transition from post-occurrence detection to a predictive safety mechanism.
• To provide critical intervention time for flight control systems.

Main Methods:

• A physics-informed multimodal deep learning framework (MTSF-Net) using a ViT-LSTM architecture.
• Fusion of seven-channel onboard sensor data (acceleration, angular velocity, vertical velocity) transformed via Continuous Wavelet Transform (CWT).
• Implementation of Calibrated Benchmark Normalization (CBN) for sensor data and a Hybrid Ordinal Loss for class imbalance.

Main Results:

• Achieved 98.26% test accuracy and 100% precursor recall for VRS.
• Eliminated fatal missed detections and false positives related to precursor states.
• Verified physical interpretability using Gradient-weighted Class Activation Mapping (Grad-CAM).

Conclusions:

• The proposed framework offers a reliable and interpretable early warning system for VRS.
• Enables proactive intervention, significantly enhancing the safety of dual-rotor UAV operations.
• Establishes a foundation for advanced, interpretable intelligent aviation safety systems.