Phy-Ture: A Physics-Guided Fusion and FeaTure Processing Framework for Unmanned Helicopter Dynamics Modeling
TL;DR
In this framework, the sequential dynamics estimates from the first-principles based model are fused with raw state-control input sequences to implement physics-guided feature augmentation and a time-series decomposition and prediction module is introduced to explicitly decouple and model dynamic components across distinct time scales, enabling the precise prediction and compensation of dynamic residuals.
Abstract
Accurate dynamics modeling for unmanned helicopters remains challenging because rotor aerodynamics, structural vibrations, and actuator dynamics are strongly coupled, nonlinear, and rapidly time varying during flight. To address this challenge, we propose a physics-guided fusion and feature processing framework for unmanned helicopter dynamics modeling (Phy-Ture). In this framework, the sequential dynamics estimates from the first-principles based model are fused with raw state-control input sequences to implement physics-guided feature augmentation. Subsequently, a multiscale temporal feature enhancement module is incorporated to reinforce the framework’s representational capability for both the short-term transient behaviors and long-term evolutionary trends of the unmanned helicopter. Furthermore, a timeseries decomposition and prediction module is introduced to explicitly decouple and model dynamic components across distinct time scales, enabling the precise prediction and compensation of dynamic residuals. Experiments on the Stanford Autonomous Helicopter Project dataset show that Phy-Ture reduces the mean RMSE by approximately 48% relative to the best-performing compared baseline under the benchmark protocol and by approximately 56% on held-out maneuver classes under the maneuver-class-based 10-fold evaluation. These results support the effectiveness of combining a structured physical baseline with adaptive temporal feature learning within the evaluated dataset and maneuver classes.