Wind-Tunnel Validation of a Reinforcement Learning Based Gust Load Alleviation Controller for a Flexible Wing

Dies ist die aktuellste Version dieses Eintrags.

Kurzfassung

This paper addresses the experimental validation of a reinforcement learning (RL)-based gust load alleviation (GLA) controller for a flexible aeroelastic wing. A soft actor critic (SAC) controller is trained offline using a linear aeroservoelastic (ASE) model at 30 m/s with broadband chirp gust excitation and is deployed as a fixed deterministic neural network policy. The policy maps outboard acceleration and flap deflection measurements to trailing edge flap commands in real time. Before wind-tunnel testing, the trained policy is evaluated in simulation using experimentally identified actuator and sensor effects, including delay, position and rate limits, backlash, and measurement noise. Wind tunnel experiments are performed in the DNW-NWB wind tunnel under harmonic gust excitation generated by rotating slotted cylinder gust generators. For the single actuator Flap 4 configuration, the controller reduces the peak-to-peak wing root bending moment (WRBM) by approximately 65% at the first bending-mode frequency. When the gust excitation is increased using two generators, the same fixed Flap 4 policy maintains a 51% reduction. These results show that an SAC policy trained on a linear ASE model can be transferred to wind tunnel hardware for targeted GLA when supported by simulation based verification and validation (V&V).

Verfügbare Versionen dieses Eintrags

• Wind-Tunnel Validation of a Reinforcement Learning Based Gust Load Alleviation Controller for a Flexible Wing. (deposited 06 Jul 2026 13:14) [Gegenwärtig angezeigt]