Optimizing Model-Based Diagnosis Complexity for Analogue Linear Systems

Fault diagnosis of analogue systems is a challeng-ing task and no fully automated solution exists. We address the task of diagnosing hard faults in linear analogue systems. We develop an algo-rithm that dynamically reduces the size of the sim-ulation model during diagnostics inference. We compare our simulation approach to the approach used by industry and academia and empirically demonstrate that our algorithm provides signifi-cant speedups on a benchmark of analogue cir-cuits with regular and semi-random topologies. We apply our algorithm to the model of a real-world satellite and demonstrate 40 times speedup. We measure a significant performance increase (200 times) for a configuration of the diagnostic search that achieves higher diagnostic accuracy, optimizing a trade-off in computational complex-ity versus diagnostic accuracy. Our framework can be directly applied to any system for which the bond graph energy formulation applies and the model optimization part of the algorithm can be used to improve the computational complexity and numerical stability of a large class of numer-ical solvers that require steady-state simulation as a part of their convergence process.