Dynamic Bayesian networks for machine diagnostics: hierarchical hidden Markov models vs. competitive learning

The failure mechanisms of mechanical systems usually involve several degraded health states. Tracking the health state of a machine, even if the machine is working properly, is very critical for detecting, identifying, and localizing the failure (i.e., diagnosis) and estimating the remaining-useful-life of the component/machine (i.e., prognosis) for carrying out proper maintenance. Hidden Markov models (HMM) present us an opportunity to estimate these unobservable health states using observable sensor signals. Hierarchical HMM is composed of sub-HMMs in a hierarchical fashion, providing functionality beyond a HMM for modeling complex systems. Implementation of HMM based models as dynamic Bayesian networks (DBN) facilitates compact representation as well as additional flexibility with regard to model structure. Regular and hierarchical HMMs are employed here to estimate on-line the health state of drill-bits as they deteriorate with use on a CNC drilling machine. In the case of regular HMMs, each HMM (that is part of a committee) competes to represent a distinct health state and learns through competitive learning. In the case of hierarchical HMMs, health states are represented as distinct nodes in the top of the hierarchy. Detailed results from regular and hierarchical HMMs are very promising and are reported in this paper.