Aircraft Anti-skid Braking Control Based on Reinforcement Learning

Abstract

Reliability and efficiency under complex working conditions are the ultimate goal of aircraft wheel brake system. In order to ensure the safety under different working conditions, traditional control methods for aircraft braking tend to be conservative in parameter adjustment, resulting in loss of efficiency. In this article, an antiskid brake control algorithm based on reinforcement learning (RL) is designed, which takes both safety and efficiency into consideration. Meanwhile, to solve the problems of generalization difficulty of different working conditions, designing difficulty of the reward function and the high failure rate in training brought by direct application of RL, we adopt task decomposition, dimension reduction of state features, action discretization and inverse RL (IRL) strategies. Finally, to avoid human–computer interaction disputes, we design a strategy in which the RL agent is pilot-activated without changing the existing hardware. The efficiency and robustness of the algorithm are proved in simulation under three typical disturbance conditions. Besides, the ground inertial bench test results show that the average deceleration rate of the proposed algorithm is 25% higher than that of the traditional algorithm, which is of engineering application value.