Model-based diagnosis deals with the problem of diagnosing faults in systems using a model of the system for guidance. This problem is complicated by the presence of hybrid dynamics in the system (continuous evolution of the system interspersed with discrete events like commands to change configuration), as well as uncertainties in the form of model approximations and sensor noise. Several model-based technologies have been developed and successfully demonstrated using discrete abstractions of the system as models. These techniques are severely restricted in model expressiveness due to the discrete nature of the models. Moreover, sophisticated model abstraction techniques, as well as algorithms to convert continuous data to discrete form, need to be developed for such an approach to work. Recently, there have been efforts to develop diagnostic engines for hybrid and stochastic systems. However, these techniques have either focused on parametric faults, or use a probabilistic approach to fault identification. Consistency-based approaches that have been successfully demonstrated using discrete models have not been extended to work with stochastic and hybrid models.
HyDE is a model-based diagnostic engine capable of detecting and isolating discrete (possibly multiple) faults in physical systems. HyDE takes as input a model of the system to be diagnosed and the telemetry/data from the actual system or from simulations of the system, and diagnoses the health state of all components in the system. The current version of HyDE has been implemented in C++.
A model of the system is developed describing the structural, transitional, and behavioral properties of the system under nominal and fault conditions. These are expressed as models of components of the system and the interconnections between the components. Data from the system, in the form of commands to the system and the sensed observations from the system, is used in conjunction with this model for fault detection and isolation. HyDE uses a combination of a consistency-based approach and stochastic approaches in which the model is used to predict the expected behavior of the system, which is then compared against the data from the system to check for consistencies. Any inconsistencies drive a search process for possible candidates that can eliminate inconsistencies. The key innovation in HyDE is the ability to deal with stochastic (uncertain) and hybrid (discrete and continuous) models and data.
This work was done by Lee Brownston of Ames Research Center and Sriram Narasimhan of the Regents of the University of California, Santa Cruz. This software is available for use. To request a copy, please visit https://software.nasa.gov/software/ARC-15570-1A