Computational Algorithm for Dynamic Hybrid Bayesian Network in On-line System Health Management Applications

Abstract

With the increasing complexity of today's engineering systems that contain various component dependencies and degradation behaviors, there has been increasing interest in on-line System Health Management (SHM) capability to continuously monitor (via sensors and other methods of observation) system software, and hardware components for detection and diagnostic of safety-critical systems. Bayesian Network (BN) and their extension for time-series modeling known as Dynamic Bayesian Network (DBN) have been shown by recent studies to be capable of providing a unified framework for system health diagnosis and prognosis. BN has many modeling features, such as multi-state variables, noisy gates, dependent failures, and general posterior analysis. BN also allows a compact representation of the temporal and functional dependencies among system components. However, one of the barriers to applying BN in real-world problems is limitation in adequately handle "hybrid models", which contain both discrete and continuous variables, with both static and time-dependent failure distributions.

This research presents a new modeling approach, computational algorithm, and an example application for health monitoring and learning in on-line SHM. A hybrid DBN is introduced to represent complex engineering systems with underlying physics of failure by modeling a theoretical or empirical degradation model with continuous variables. The methodology is designed to be flexible and intuitive, and scalable from small, localized functionality to large complex dynamic systems. Markov Chain Monte Carlo (MCMC) inference is optimized using a pre-computation strategy and dynamic programming for on-line monitoring of system health. Proposed Monitoring and Anomaly Detection algorithm uses pattern recognition to improve failure detection and estimation of Remaining Useful Life (RUL). Pre-computation inference database enables efficient on-line learning and maintenance decision-making. The scope of this research includes a new modeling approach, computation algorithm, and an example application for on-line SHM.