A generative machine learning framework for anomaly response in cyclical processes in the ECLSS on a deep space habitat
Michael Ibrahim, Samuel P. Eshima, Nagi Gebraeel, James Nabity, Stephen K. Robinson · Acta Astronautica · 2025
In recent years, there has been growing interest in the development of deep space habitats (DSHs) for Lunar and Martian exploration missions. While such habitats may rely on many existing spacecraft technologies in their development, they are faced with unique challenges that require careful consideration. A key example of such challenges is the unusually long communication delays from Earth, which could be on the order of three to twenty minutes. As a result, DSHs require a higher level of Earth-independence than existing spacecrafts when handling and performing various critical tasks and activities. Another key challenge is the lack of historical data that could be leveraged to train data-driven anomaly response models due to the novelty of DSHs. In this work, we study the critical task of anomaly response in the Environmental Control and Life Support System (ECLSS) on a DSH while minimizing Earth-dependence, with a focus on cyclical processes. More specifically, we develop an autonomous machine learning (ML) framework capable of detecting and diagnosing (unknown and known) operating conditions in the ECLSS, respectively. Our proposed framework maintains a growing registry of known operating conditions (i.e. baselines and faults), whereby a human subject matter expert (SME) is prompted to label any unknown operating conditions that occur, triggering their automatic addition to the registry. The framework is then capable of autonomously triggering a retraining, as well as generating a training dataset utilizing generative ML capabilities in order to learn the new operating condition. As a case study, we focus on the CO 2 removal system to validate our proposed framework. To that end, we utilize experimental data generated from a CO 2 removal system testbed, where we carefully examine all aspects of our proposed framework and validate its capabilities.