Predicting Dynamic Behavior Using Physical System Models with Algebraic Loops
Pieter J. Mosterman
Institute for Robotics and System Dynamics
DLR Oberpfaffenhofen
Eric-Jan Manders
Department of Electrical and Computer Engineering
Vanderbilt University
Gautam Biswas
Knowledge Systems Lab
Stanford University
Abstract
Our system for monitoring and diagnosis of abrupt faults in complex dynamic systems,
TRANSCEND, relies on system models to predict dynamic
behavior in response to abrupt faults. The use of qualitative predictions of this transient behavior mitigates complexity issues and
convergence problems that exist in numerical diagnosis approaches.
After a fault is detected, future behavior for all possible causes is
predicted and captured in the form of a signature. Progressive
monitoring uses these signatures to validate hypothesized causes and
prune the set of candidates. In this framework, it is important to model
the actual system at a level of detail that relates to the measurement
bandwidth. Too much detail results in fast behaviors that are modeled
as continuous transients but that appear as discontinuous changes in
the measured signals. In many cases, abstracting these small parameter
values away may result in algebraic dependencies between variables,
i.e., variables affect one another instantaneously without integrating,
state, behavior. Because of the compensating effect of physical system
behavior, a straightforward deviation propagation results in predictions
that are unknown in a qualitative sense. This paper describes an
algorithm that recognizes such dependencies and propagates
compensating effects without introducing unnecessary conflicts. The
effectiveness of the approach is demonstrated by diagnosing a
punctured hose in the cooling system of a combustion engine.