Prognostics for Autonomous Deep-Space Habitat Health Management under Multiple Unknown Failure Modes

Deep-space habitats (DSHs) are safety-critical systems that must operate autonomously for long periods, often beyond the reach of ground-based maintenance or expert intervention. Monitoring health and anticipating failures are essential for safe operations. Prognostics based on remaining useful life (RUL) prediction support this goal by estimating how long a subsystem can operate before failure. Critical DSH subsystems, including environmental control and life support, power generation, and thermal control, are monitored by many sensors and can degrade through multiple failure modes. In practice, these failure modes are often unknown, and the sensors providing useful information may vary across modes, making accurate RUL prediction challenging when failure data are unlabeled. We propose an unsupervised prognostics framework for RUL prediction that jointly identifies latent failure modes and selects informative sensors using unlabeled run-to-failure data. The framework has two phases: offline sensor selection and failure mode identification, and online diagnosis and RUL prediction. In the offline phase, failure times are modeled using a mixture of Gaussian regressions, and an Expectation-Maximization algorithm simultaneously clusters degradation trajectories and selects mode-specific sensors. In the online phase, low-dimensional features from selected sensors diagnose the active failure mode and predict RUL through a weighted functional regression model. The framework is evaluated on a simulated dataset capturing key telemetry challenges in DSH systems and on the NASA C-MAPSS benchmark. Results show improved prediction accuracy and clearer identification of informative sensors and failure modes than existing methods.

💡 Research Summary

The paper addresses the critical need for autonomous health management and remaining useful life (RUL) prediction in deep‑space habitats (DSHs), which operate for extended periods without ground support and are monitored by thousands of sensors. Because DSH subsystems can fail through multiple, unknown failure modes and the informative sensors differ across modes, traditional supervised prognostics that rely on labeled failure data are infeasible. The authors propose a fully unsupervised prognostics framework that jointly discovers latent failure modes and selects mode‑specific sensors, then uses these selections for real‑time diagnosis and RUL estimation.

In the offline phase, the authors model failure times (time‑to‑failure, TTF) with a mixture of Gaussian regressions (MGR). An Expectation‑Maximization (EM) algorithm is devised to simultaneously (i) cluster each run‑to‑failure trajectory into a latent failure mode and (ii) perform sensor selection for each mode using an Adaptive Sparse Group Lasso (ASGL) penalty on the regression coefficients. This yields, for every discovered mode, a compact set of sensors that carry the most predictive information.

The online phase begins by extracting low‑dimensional functional features from the selected sensors using Multivariate Functional Principal Component Analysis (MFPCA). These features feed a K‑Nearest‑Neighbor classifier that identifies the currently active failure mode. Once the mode is diagnosed, a weighted functional regression model—where recent observations receive higher weights—is applied to map the functional sensor trajectories to the remaining time before failure, producing the RUL estimate.

The framework is evaluated on two datasets. The first is a high‑fidelity simulation of DSH telemetry that mimics the challenges of massive sensor counts, variable signal‑to‑noise ratios, and unlabeled multi‑mode failures. The second is the NASA Commercial Modular Aero‑Propulsion System Simulation (C‑MAPSS) turbofan engine benchmark, a widely used prognostics testbed with similar sensor‑fault characteristics. Across both datasets, the proposed method outperforms baseline approaches, including single‑mode regression, PCA‑Lasso sensor selection, and recent deep‑learning multi‑mode models, achieving lower mean absolute error and higher R² scores. Moreover, the sensors identified by the algorithm align with domain‑expert knowledge of failure mechanisms, enhancing interpretability.

Key contributions are: (1) an unsupervised EM‑based mixture‑of‑regressions framework that discovers failure modes and performs mode‑specific sensor selection without any expert labeling; (2) a practical online pipeline that combines MFPCA, nearest‑neighbor mode diagnosis, and weighted functional regression for real‑time RUL prediction; (3) extensive validation on both simulated DSH data and a real aerospace benchmark, demonstrating superior accuracy and explainability. Limitations include the need to pre‑specify the number of latent modes, sensitivity to EM initialization, and computational cost for very large sensor suites. Future work is suggested on Bayesian non‑parametric mixtures to infer the number of modes automatically, online EM variants for faster adaptation, and hybrid physics‑data models to further improve robustness under the extreme conditions of deep‑space missions.