Bond graphs [21] provide a systematic framework for building consistent and well constrained models of dynamic physical systems across multiple domains with inherent causality constraints that provide effective and efficient mechanisms for diagnosis. An added advantage of the bond graph derived representation is their direct applicability to qualitative processing, which makes them applicable in situations where precise numerical information may not be available. However, analytic system models derived from bond graphs are also amenable to quantitative simulation and analysis.
In our work, a causal dependency graph is derived from a bond graph to provide the system model. Links among system parameters and variables in a causal dependency graph are extended by temporal properties. Propagating effects of deviant observations to hypothesized causes (i.e., faults) can now be classified as instantaneous versus those that have delayed effects. Delayed effects can be further classified by the order of the effect, e.g., first order, second order, etc. The causal temporal models are derived from a bond graph model that adequately captures the dynamic characteristics of system behavior.
Abrupt faults cause dynamic system behavior and the resulting transients take system behavior from its nominal steady state of operation to a new steady state. Based on a model of the system dynamics, these transients can be effectively and efficiently applied to quickly isolate root-causes for deviating behavior. Magnitudes, slopes, and discontinuous changes at the time of failure for individual observations can be applied in a qualitative reasoning framework. Furthermore, the new steady state that the system achieves can be used as a final mechanism for fault isolation. However, often it is hard to determine whether a system has reached a new steady state. Therefore, future work will focus on using the steady state graph for diagnosis of incipient faults, i.e., those that do not cause dynamic behavior.