Reinforcement Learned Adversarial Agent for Active Fault Detection in Space Habitats

NASA intends on building a U.S.-led physical base on the moon capable of supporting human life. This ldquo;Sustainability Baserdquo; will require habitat environmental control and life sustainment systems. Such systems are necessarily complex, the volumes of sensor data are large and not well-suited for human-only monitoring, and the consequences of system failure are severe. Thus, to sustain optimal performance and avoid catastrophic failures, NASA seeks a health management system that will continuously monitor and quickly and accurately diagnose faulty system behavior.Navatek proposes to develop a fault prediction and detection solution that improves NASArsquo;s ability to reveal latent, unknown conditions while also improving its detection time and reducing the rate of false positive and negative detections of known conditions that would lead to failure of the life sustainment system. Our approach feeds historical and real-time sensor data to a digital twin of the life sustainment systems, which is a digital simulation of the entire functioning system and its environment. This digital twin is used by a reinforcement learning adversarial agent to simulate many possible scenarios into the future. The adversarial agent autonomously learns the environmental and system perturbations that lead to faults in its simulations, thus providing a method for prediction. These predictions are continuously compared against new incoming data to detect faults and further improve the digital twinrsquo;s accuracy.If successful, our proposed solution will provide NASA with an early warning system for faults in the life sustainment systems on space habitats, particularly integrity of the structural and HVAC systems. We will also show how our digital twin and reinforcement learning adversarial agent approach can be generalized to monitor other space habitat systems.nbsp;