Model-Driven Deep Graph Learning for Risk-Aware Anomaly Response

Future deep space habits will require resilient and autonomous operations that have a holistic understanding of the systems contained within. The HOME team stated that a fundamentally disruptive approach is required to safely perform human-exploration driven by a holistic perspective on systems across a wide range of domains. Our approach addresses the critical area of balancing human and autonomous system intervention by leveraging the capabilities inherent in model-based systems engineering (MBSE) to understand the complexity of a fault or anomaly to the mission and the safety of the astronauts. By having a systems perspective on the FDIR, a more cohesive story can be presented to astronauts, including potentiality of cascading failures, recommended courses of action, time to failure predictions, and other human factors in risk analysis. To address the need for holistic understanding, we leverage an ensemble approach that combines MBSE, graphical models, probabilistic graphical models, semantic learning, and decision theory to provide anomaly response and fault action impacts. Our innovative approach disaggregates the problem into three components: 1. Model-based systems engineering and semantic understanding that gathers all the design information, model parameters, interrelations from the systems' models and expands their context with semantic understanding of the taxonomies and classifications of components. 2. Online Guided Deep Structure Learning provides the fault propagation modeling of the entire system that is guided by a priori information about the system from the system model but learns the real characteristics of the sensors and their interdependencies through physics-informed online learning. 3. Risk-aware Anomaly Response is the final module that makes decisions based upon actual relationships and an understanding of the impact a decision will have across the entire system, the mission, and any safety concerns.