Aerodynamic stall recovery using artificial intelligence techniques

Abstract

According to International Civil Aviation Organization (ICAO) in 2019 just before Covid-19 started, 4.5 billion people worldwide travelled by plane. That is nearly half the population of the world. Although, accidents happen, statistically, it is the safest mode to travel from one place to another as of now. This is because of extensive training given to pilots, the on-board automation and protection systems of modern commercial aircraft. Despite the on-board automation and protection systems, aerodynamic stall is still a possible occurrence and pilots undergo stall detection and recovery training to deal with such scenarios. Nevertheless, accidents have occurred due to pilot error during stall recovery. This work uses combination of Reinforcement Learning (RL) algorithm, Behavioral Cloning (BC) and Deep Learning (DeL) based regression model to train multiple Machine Learning (ML) models to automatically recover an aircraft from a wings level (1G) stall, stall during a turn and stabilize it. The RL environment consists of X-Plane flight simulator, NASA’s XPlaneConnect plugin to interface X-Plane with python programming language. The design of whole setup and implementation of the algorithms is discussed, together with the training and testing of the ML models in this dissertation. The stall recovery process was divided into two parts. The first part was about reducing the Angle of Attack (AoA) below the critical AoA. Once, the current AoA is less than critical AoA then the side stick controls are handed over to the second ML model. The first agent is based on Deep Deterministic Policy Gradient (DDPG) algorithm, which has been pre-trained using BC technique. The data which is used to pre-train the actor network of DDPG is recorded with the help of expert pilots. Pre-training the actor network helps the RL agent to converge at a faster rate and learn a policy which is similar to that of an expert pilot. The pre-training of the actor network significantly reduced the time taken to train the first agent to recover from a stall as the actor network performed like an expert pilot. Once, the agent learns a policy which is similar to expert’s policy. A random noise is added to the output of the network to help the agent to explore the RL environment. This exploration is needed to help the agent to find better policies to recover from an aerodynamic stall. The second agent is a DeL regression model, which is trained on a different expert recorded dataset. The DL regression model is responsible to stabilize the aircraft and reach a safe airspeed once the RL model hands over the side stick control to the regression model. The results obtained in this work are satisfactory as the ML models have been able to recover from stall events at various altitudes between 3,000 feet and 30,000 feet. This dissertation present and discuss the training and test results in further detail and will examine the sensitivity of the algorithms to various other factors.