NASA Logo, National Aeronautics and Space Administration

General Diagnostic Issues

There are two key issues that make the diagnosis task hard, in addition to complexity introduced by the stochasticity and hybrid nature of the systems.

  1. Limited Observability – The number of sensors providing observability into the system is limited for several reasons (such as cost of sensors, inaccessible locations, limited bandwidth, etc.). As a result, faults are typically not directly observable and have to be isolated by reasoning about the system using the available (limited) sensor information.
  2. Time-delayed Symptoms – The manifestation of the effects of the faults (symptoms) is typically not seen immediately (at the same instant as the fault occurrence) for a variety of reasons. As a result, the process to isolate faults has to reason backwards and forwards in time. Some possible reasons for the time delay in manifestation of fault effects are:
    1. Integrating effects due to the presence of components with state, such as capacitors and storage tanks.
    2. Observable variables may be affected by the faults only in certain discrete mode configurations.
    3. Presence of noise and uncertainty make it impractical to base decisions on a single point of comparison.

Model-based Diagnosis (MBD)

Model-based Diagnosis (MBD) approaches provide several advantages over other techniques for diagnosis:

  • MBD separates modeling and reasoning by applying the same reasoning algorithms to models of different applications and so only models need to be built for new application
  • MBD does not require the enumeration of faults and externally visible manifestations of these faults since the model compactly captures structure and behavior of the system without enumeration.
  • Models already built for other purposes like design, simulation etc. may be reused for MBD.
  • Models are modular, compositional and hierarchical in nature which allows for reuse of models in different applications via model libraries

Stochastic Hybrid Systems

Most real world systems that need to be diagnosed tend to exhibit stochastic and hybrid properties. The stochasticity arises from several sources including insufficient/inaccurate information about the functioning of the system, sensor and other noise sources, unknown operating environments, approximate models among others. The hybrid nature arises from the switching nature of components either through commands (e.g. OPEN Valve) or autonomously (Fuse arising from high current). So an MBD engine has to take into account both of these properties.

MBD of Stochastic Hybrid Systems

Approaches to diagnosis of stochastic hybrid systems either consider a highly abstracted model of the system (Livingstone 2) or deal with single parametric faults (TRANSCEND) or take a purely probabilistic view (Dearden et al., Hofbaur and Williams).

In HyDE, we try to bring together all these approaches and provide a framework that allows the user to model in different paradigms and use different algorithmic approaches for diagnostic reasoning.

First Gov logo
NASA Logo - nasa.gov