Structural abstraction for model-based diagnosis with a strong fault model

摘要

Model-based diagnosis (MBD) is a fundamental approach for automated diagnosis, in which a model of the diagnosed system is used to identify abnormally behaving system components. As systems become large-scale and more complex, their models also grow in size, and consequently applying MBD becomes more computationally challenging. Structural abstraction was shown to be effective in scaling up MBD algorithms to larger systems. However, past work on using structural abstraction in MBD assumed, either explicitly or implicitly, a weak fault model (WFM), i.e., that the system model specify only the normal behavior of the system components. Therefore, the resulting diagnoses can be inconsistent with existing knowledge about how the system behaves when it is abnormal. System models that contain such information are said to have a strong fault model (SFM). In this work, we show that a standard approach for using cones abstraction, a form of structural abstraction that was shown to be useful for directional systems, does not work for systems with a SFM. Then, we propose several sound and complete algorithms that can use a cones abstraction effectively to diagnose systems with a SFM. Some of these algorithms use Machine Learning techniques to predict which cones will not be useful in the diagnosis process and should be discarded. Empirical evaluation on benchmark systems that model Boolean circuits shows that our algorithms are very effective in practice. The empirical evaluation also sheds light on how various system properties affect the comparative performance of the proposed algorithm.