Adaptive AI Hybrid Control Method for Large Unmanned Aerial Vehicle Systems in Dynamic Environments

In this research paper, we propose a novel adaptive AI hybrid control architecture designed to overcome the model uncertainty and time-varying noise characteristics encountered in large unmanned aerial vehicle (UAV) systems. Traditional filtering techniques with fixed parameters struggle to maintain accuracy under dynamic vibration, mechanical wear, and environmental fluctuations, while pure AI-based approaches face practical challenges due to real-time computational demands and the lack of clean ground-truth IMU data. To address these limitations, our framework integrates a low-level control layer—responsible for rapid stabilization via adaptive Kalman filtering and adaptive PID control—with a high-level supervisory layer that employs an LSTM-based residual learning model to perceive evolving noise dynamics and optimize control parameters in real time. Experimental evaluations under realistic multi-phase dynamic vibration scenarios demonstrate that the proposed Residual-LSTM achieves 49.6–51.2% RMSE reduction across the three gyroscope axes relative to the raw noisy IMU signal and over 75% RMSE reduction in highly dynamic vibration phases relative to the fixed classical filter, highlighting enhanced adaptability and robustness in non-stationary operating conditions.