Using a Dynamic Bayes Network to Represent System State for Fault Diagnosis

On space missions, astronauts rely on Mission Control for help in resolving emergency situations. However, for longer-duration or further distance missions, communication delays with Mission Control can significantly impact time-sensitive emergencies. To address this, there needs to be an on-board system that can help the crew diagnose faults [Marquez et al., 2017]. Human-human teams such as Mission Control and astronauts use shared mental models to communicate and coordinate more efficiently in time-critical situations. Human-human teams also bring creativity for robust responses to fault diagnoses [Cannon-Bowers and Salas, 1990, Rerup, 2001]. An AI system on board can take the place of Mission Control in emergencies, but needs to be able to understand the human’s beliefs for effective communication and creativity similar to human-human teams [Scheutz et al., 2017]. This project proposes a Dynamic Bayes Network to monitor the system over time while communicating with the human agent for observations of the system state. This model estimates the state of the system based on the human’s understanding of the system’s functioning to attempt to estimate the human’s mental model of the system for effective human-AI teaming. This project evaluates different levels of observability in the system to determine the effect of the number of observations provided by the human-agent on system monitoring. The experiments also simulate a fault situation in order to explore how faulty observations impact the system’s estimate. The results show that a larger number of observations over different parameters provide the best picture of the system. This model also shows potential for fault diagnosis as it is able to recognize that the faulty observations are extremely unlikely.