Deep reinforcement learning-based ditching vertical velocity control for a multi-fault wing-in-ground craft

Abstract

When a wing-in-ground craft (WIG) ditches, the water impact load may not only cause structural and equipment damage but also endanger the safety of passengers. Nevertheless, the impact can be considerably lessened by decreasing the WIG’s vertical velocity upon entry into the water. Current research on forced landing automatic control focuses primarily on land vehicles, while research on WIG ditching vertical velocity control is scarce. However, the path-following method applied for land emergency landings is inapplicable to ditching for a multi-fault WIG. When multiple malfunctions (engine, maneuvering mechanism, and sensors) occur simultaneously, it is difficult for a WIG to follow a planned path. This paper proposes a deep reinforcement learning-based ditching vertical velocity control method for multi-fault WIGs. The method seeks to reduce the vertical water entrance velocity during WIG ditching by optimizing the control strategy while only the elevator and height sensors are operational. A dense reward function was developed based on a sparse reward function that emphasized the completion of the ditching task and the vertical velocity at the moment of termination. Penalties are employed to restrict the WIG’s pitch angle to a safe range and prevent excessive landing time. Without following the artificially established target path, the developed reward function can directly optimize the vertical velocity control strategy during the WIG ditching process. The effectiveness of the proposed method is verified by simulations in which the WIG ditches at three various initial flight altitudes. In comparison to the path-following method for emergency landings of land vehicles, the proposed method reduces the WIG’s vertical water entrance velocity by more than 80% in all conditions. In addition, the effect of reward function hyperparameters on the optimization of the control strategy is investigated. The proposed method is instructive for multi-fault WIGs ditching vertical velocity control.