Problem

Aircraft upset conditions, such as stalls and spins, are a major cause of aviation accidents. Traditional Pilot Activated Recovery Systems (PARS) rely on classical control methods, which often involve labor-intensive hand-crafted designs and may not achieve optimal or comprehensive recovery maneuvers. These methods are frequently constrained, preventing them from exploring non-intuitive but highly effective solutions across the entire flight envelope. There is a need for a more efficient and adaptive recovery system that can enhance flight safety and reduce the complexity of control engineering.

Method

The researchers developed an AI-driven PARS model utilizing a modern reinforcement learning (RL) architecture, specifically the Soft Actor-Critic (SAC) algorithm. The system was trained using a high-fidelity flight dynamics model of an advanced jet trainer in a simulation environment. A key aspect of the methodology was the iterative design of the reward function, which started with targets for zero roll (ϕ) and gamma (γ) angles. To address issues like action oscillation, penalties for command derivatives were introduced. Furthermore, a critical feature was the incorporation of negative-g force punishments, leading to episode termination if g-forces dropped below -2g, to ensure pilot safety. Hyper-parameter optimization for the SAC algorithm was performed using Optuna, an efficient search method that leverages previous trials. The state space was minimized for complexity reduction, and the action space was limited to elevator and aileron stick commands, based on expert feedback.

Results

The final RL-based PARS model demonstrated significant performance improvements over classical control methods. Evaluations by domain experts and comparisons using simulation data showed that the AI model achieved faster recovery times for both roll (ϕ) and gamma (γ) angles. For instance, in one case (initial ϕ=-100, γ=45), the AI recovered ϕ in about 6 seconds compared to 8 seconds for the classical controller, and fully recovered γ in 8 seconds, whereas the classical controller could not achieve full recovery. Crucially, the AI model successfully avoided negative-g exposure for the pilot, which was a recurring issue with classical methods. This was achieved by learning an inverse flight maneuver to recover gamma while maintaining safe g-forces, then recovering the roll angle. These results validate the significance of AI in avionics and suggest that such systems can surpass traditional control mechanisms.

Implications

This research validates the potential of integrating state-of-the-art reinforcement learning with domain-specific aerospace engineering knowledge to create highly effective flight control systems. The development of an unconstrained PARS controller that outperforms traditional methods in recovery speed and pilot safety marks a significant advancement for avionics. The ability of the RL model to cover a wider upset condition space and generate optimal, non-intuitive recovery maneuvers without hard constraints opens new avenues for enhancing flight safety and operational efficiency in advanced aircraft. The findings lay a foundation for further research in applying AI to complex aerospace engineering challenges, potentially leading to more robust and adaptive autonomous flight systems.