A simulation framework for life-support anomaly response in deep-space exploration habitats

 The increasing autonomy required for Environmental Control and Life Support Systems  (ECLSS) in human spaceflights, especially for missions beyond low Earth orbit, necessitates  advanced self-awareness and self-sufficiency onboard spacecraft. This autonomy is supported  by developing and integrating novel flight procedures that rely heavily on onboard informa tion technologies, such as computational models, rather than ground control resources. This  research focuses on extending a Probabilistic Graphical Model (PGM) digital twin framework  to integrate these disparate models and serve as tools for onboard query processing to sup port anomaly response during operation. Specifically, we use directed PGMs, i.e., Bayesian  Networks (BNs), which are Machine Learning models described by a network where nodes  are random variables and arrows represent their causal relationships. Two categories of  challenges are considered: (1) analysis tasks must be transferred from ground support to be  performed autonomously onboard space habitats, and (2) the modeling uncertainties pertain ing to onboard computations must be quantified to provide insight to the users. Specifically,  the research questions addressed in this work are as follows:  

1. What information construct can leverage existing modeling standards and libraries to  formalize a computer-readable characterization of system states and model parameters  for probabilistic calibration execution and assessment in space habitat operations?  
2. What design rules are to prioritize for a PGM-based multi-model calibration maximiz ing parameter assignment ability and minimizing computational cost given resolution  discrepancies between diagnosis labels, expert knowledge, and parameter space?  
3.  What selection rules can leverage both domain knowledge and statistical heuristics to  algorithmically prune a PGM for multi-model calibration thereby maximizing param eter assignment ability and minimizing computational cost?  

First, this work reviews existing information technologies for autonomous operations  of ECLSS. We propose a framework to migrate calibration procedures, employed by mission  control for computer models, from the ground to space habitats during operation. This  vision consists of deploying models in a digital twin providing an integrated environment  for a supervision agent to easily access modeling resources through a software interface.  Specifically, this work suggests calibrating simulation models by encoding domain knowledge  into a probabilistic mathematical abstraction of subsystem states and model parameters.  

The research further investigates the application of PGMs to quantify the uncertain ties pertaining to the estimation of system states in ECLSS. PGMs are utilized to integrate  expert knowledge with data, thus improving accuracy in diverse operational scenarios, such  as temperature prognosis. A detailed guideline for implementing PGM-based digital twins is  provided, emphasizing the iterative process of outlining, designing, calibrating, and evaluat ing the model. This guideline is validated through experiments simulating ECLSS behavior  in an analogous terrestrial testbed, providing insights into decision-making problems based  on habitability criteria.  

A challenge addressed in this research is the migration of simulation models from  mission control to space habitats. This involves automating the calibration of heterogeneous  models, traditionally managed by subsystem specialists on the ground. PGMs are used to  encode dependencies between simulation models to facilitate integrated calibration onboard.  This approach is demonstrated in a docking scenario involving a Carbon Dioxide (CO2)  removal fault, integrating machine learning, physics-based, and knowledge-based models to  infer the most likely simulation parameters based on diagnosed system states. The following  research gaps are identified: (1) parameter and system description, and (2) PGM design for  multi-model calibration. These gaps were further investigated in two research studies.  

The first study proposes an ontology for simulation parameters and system states that  extend existing standards to include attributes for autonomous information retrieval. This  ontology enables autonomy by providing a formal representation of attributes and relations  necessary to retrieve information without expert intervention in deep space, including pa rameter attributes (e.g., activation status and probabilities), and sub-system properties (e.g.,  hierarchical topology and power dependencies).  

In the second study, we investigated how a PGM can be designed to capture system  and parameter dependencies, enabling decision-makers to conduct “what-if” analyses. The  proposed design process leverages Model-Based System Engineering (MBSE) resources and is  compared with existing design strategies. We introduce heuristics to (1) design a large PGM  capturing existing knowledge about system states and model parameters, and (2) prune the  network when necessary to limit computational costs while maximizing calibration ability.  

Overall, this research advances the development of autonomous ECLSS operations  by integrating and calibrating relevant simulation models onboard, thereby enhancing the  resilience and self-sufficiency of space habitats for future human spaceflights.