Monitoring a Complez Physical System using a Hybrid Dynamic Bayes Net

Uri Lerner, Brooks Moses, Maricia Scott, Sheila A. McIlraith, Daphne Koller

Introduction

Future Mars missions depend on the ability to produce oxygen and fuel in situ. NASA’s Reverse Water Gas Shift (RWGS) system is a critical technology designed for this purpose, converting the Martian atmosphere into life-sustaining resources. However, operating such a system autonomously for hundreds of days requires a highly sophisticated monitoring framework capable of handling complex, non-linear physical dynamics and unreliable sensor data. This paper presents a Hybrid Dynamic Bayesian Network (DBN) approach that provides robust, real-time tracking of the RWGS system.

The Challenges of Monitoring RWGS

Monitoring the RWGS system is a significant challenge for traditional control and estimation theories. The system exhibits three primary difficulties:

  1. Nonlinearity: The underlying physics of gas flows and chemical reactions are inherently non-linear.
  2. Sensor Imperfections: Sensors are not only noisy but also prone to time-varying biases that depend on the system's operational state.
  3. Stiff Dynamics: The system involves "stiff" temporal disparities, where pressure waves stabilize in milliseconds, while gas composition changes occur over hours.
Schematic of the RWGS system alongside sample sensor data for CO flow and pressure differentials.

Traditional methods like the Extended Kalman Filter (EKF) often fail due to linearization errors, while Particle Filters can suffer from "weight collapse" in high-dimensional spaces, where a single sample dominates the probability mass.

Methodology: Hybrid Dynamic Bayesian Networks

The researchers modeled the RWGS using a Hybrid DBN consisting of 33 discrete and 184 continuous variables. To overcome the limitations of standard filters, the authors introduced several technical innovations:

Numerical Integration for Nonlinearity

Rather than linearizing the system (as in EKF), the authors treated the tracking problem as a numerical integration task. They employed Exact Monomial rules to approximate the probability distribution. This framework generalizes the Unscented Filter (UF), which is shown to be a specific case (precision 3) of this broader integration approach.

Handling Multi-Scale Dynamics

To manage the computational burden of micro-second pressure fluctuations within a one-second observation window, the authors used fixed-point equations. Near-instantaneous processes are modeled as simultaneous equations that converge to a steady state within a single time step, avoiding the need for extremely fine-grained temporal simulation.

Dynamic Bias Modeling

To account for sensor drift, hidden "bias" variables were integrated into the state. A contraction factor ($\gamma < 1$) was applied to pull these biases toward zero. This prevents the model from incorrectly attributing real system faults to sensor bias unless there is persistent, strong evidence of a discrepancy.

Comparison of filtering accuracy: Higher-order numerical integration rules (Precision 5) significantly outperform the EKF.

Key Results and Performance

The Hybrid DBN approach demonstrated significant advantages over existing methods:

  • Efficiency: The algorithm processed each time step in approximately 20ms on 2002-era hardware, making it roughly 750x faster than standard particle filtering.
  • Reliability: In simulations, the DBN remained stable and accurate where Particle Filters (even with 10,000 samples) failed due to variance collapse.
  • Real-World Validation: Using 13,875 time steps of real-world data from a NASA prototype, the model successfully tracked complex transitions, such as the "rebound" effect in gas flows during a $CO_2$ supply shutdown.

Conclusion and Future Work

The Hybrid DBN framework provides a robust solution for monitoring complex physical systems with multi-scale dynamics and non-linearities. While the current model assumes the discrete state (e.g., valve positions) is known, future work aims to integrate Rao-Blackwellized Particle Filtering (RBPF). This extension will enable the system to simultaneously track continuous physics and discrete operational modes, allowing for real-time diagnosis of hardware failures like leaks or compressor malfunctions.

Summary updated

Abstract

The Reverse Water Gas Shift system (RWGS) is a complex physical system designed to produce oxygen from the carbon dioxide atmosphere on Mars. If sent to Mars, it would operate without human supervision, thus requiring a reliable automated system for monitoring and control. The RWGS presents many challenges typical of real-world systems, including: noisy and biased sensors, nonlinear behavior, effects that are manifested over different time granularities, and unobservability of many important quantities. In this paper we model the RWGS using a hybrid (discrete/continuous) Dynamic Bayesian Network (DBN), where the state at each time slice contains 33 discrete and 184 continuous variables. We show how the system state can be tracked using probabilistic inference over the model. We discuss how to deal with the various challenges presented by the RWGS, providing a suite of techniques that are likely to be useful in a wide range of applications. In particular, we describe a general framework for dealing with nonlinear behavior using numerical integration techniques, extending the successful Unscented Filter. We also show how to use a fixed-point computation to deal with effects that develop at different time scales, specifically rapid changes occurring during slowly changing processes. We test our model using real data collected from the RWGS, demonstrating the feasibility of hybrid DBNs for monitoring complex real-world physical systems.
BibTeX
@inproceedings{lerner2002monitoring,
  author = {Uri Lerner and Brooks Moses and Maricia Scott and Sheila A. McIlraith and Daphne Koller},
  title = {{Monitoring a Complez Physical System using a Hybrid Dynamic Bayes Net}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2002},
  url = {https://openreview.net/forum?id=yGbmVRdj7U},
}
Figure