A Reinforcement Learning Control and Fault Detection Method for the MADNI Drone

Amidst the tumultuous storms and challenging weather conditions that have engulfed Northern Ireland at the end of 2023 into 2024, highlights the demand of Unmanned Aerial Vehicles (UAVs) equipped with Search and Rescue (SAR) capabilities to revolutionise emergency response efforts and bolstering resilience in the face of adversity. This paper explores the application of Reinforcement Learning (RL) techniques, specifically Deep Deterministic Policy Gradient (DDPG) methods, for enhancing control and fault detection capabilities in the real-life Manoeuvrable Autonomous Drone for Navigation and Intelligence (MADNI). We investigate the performance of DDPG agents trained with different optimisers, including Adaptive Moment Estimation (ADAM), Root Mean Square Propagation (RMSprop), Stochastic Gradient Descent (SGD), Adaptive Gradient Algorithm (AdaGrad), and Stochastic Gradient Descent with Momentum (SGDM). Our study aims to assess the effectiveness of these optimisation methods by improving the stability, convergence speed, and fault detection accuracy of the MADNI model. By conducting comprehensive simulations and experiments, we evaluate the ability of DDPG-based RL agents to navigate and detect faults in dynamic and uncertain environments. The findings of this research contribute to advancing autonomous systems' reliability and adaptability by identifying optimal strategies for training RL agents in UAV control and fault detection tasks.