Deep reinforcement learning control for stall flutter via active camber morphing

Stall flutter suppression is complicated due to its nonlinearity and unsteadiness. A deep reinforcement learning (DRL) framework with the Deep-Q-network algorithm is established for stall flutter suppression by camber morphing. A reduced-order predictive model for stall flutter is developed for rapid data acquisition and utilized as a DRL training environment. In the fast-responsive environment, the smart controller is trained using a low-dimensional, dense neural network. Then, the trained smart controller is applied to fluid-structure-control coupled numerical simulation to validate its effectiveness and test its generalization ability. The results show that the smart controller trained under an airspeed of 8 m/s and structural stiffness of 0.3 N m/rad can completely suppress the pitch motion of the airfoil in both baseline case and generalized cases: among airspeeds of 8–11 m/s and structural stiffness of 0.1–0.3 N m/rad. Additionally, in tested cases, the trailing edge morphs at an identical frequency as pitch motion with a 0.7p phase offset. Higher control efficiencies are observed with control origins during the upstroke process of pitch. By analyzing from the flow perspective, the reduction and delay of the leading-edge vortex are found to be the main contributors to stall flutter suppression. Delays in the secondary vortex and the synchronization of the secondary vortex and trailing edge vortex also have a credible suppression effect on stall flutter.