Our work on measurement selection illustrates the trade-off between using higher order effects versus considering more measurements in diagnostic analysis. In real situations, cascading multiple faults are more likely than independent multiple faults. Cascading faults are best handled by quick analysis of transients to establish root-causes and then suspend diagnosis when other faults begin to influence the measured transients. In spite of the loss of precision, the results are more practical from a computational viewpoint. Furthermore, the measurement selection algorithm allows a theoretical assessment of the quality of the diagnosis algorithms, which forms a benchmark for simulation. This measure is hard to obtain by exhaustive testing because the resolution of fault isolation depends on the numerical values of the system parameters, the size of the parameter deviation, and the noise level. If no unique fault is isolated by the diagnosis engine, it is unknown whether this is inherent to the diagnosis algorithms or because of the system parameters. Note that a decreased diagnostic resolution is acceptable as long as the true fault is part of the hypotheses.
The current measurement selection algorithm uses a naive exhaustive search approach to pick subsets of possible measurements to achieve optimal diagnosability. In general, this search is exponential. Currently, we are looking at other graph theoretic algorithms and heuristic schemes based on greedy search to reduce the complexity.