Critical Observations in Model-Based Diagnosis

In model-based fault diagnosis, the ability to identify the key observational data that leads to system abnormalities is highly valuable. This paper introduces a framework and algorithm for identifying key observational data. The framework determines which observations are crucial for diagnosis by abstracting the raw observational data into “sub-observations.” A “key sub-observation” is defined as the most abstracted sub-observation that still leads to the same minimal diagnosis set as the original observation.

This research is a collaboration between Cody James Christopher from Australia and Alban Grastien from France, affiliated with the Data61 center at the Commonwealth Scientific and Industrial Research Organisation (CSIRO) in Australia and the French Alternative Energies and Atomic Energy Commission (CEA), respectively. Their work was published in the Artificial Intelligence journal in 2024.

The researchers first outlined the basic framework and concepts of model-based diagnosis. The framework consists of three main components: a system model, observational data, and a diagnosis hypothesis space. The system model describes all possible behaviors of the system; observations are the perceptions of the actual system behavior, which may come from sensor readings or log records; and diagnosis hypotheses correspond to potential fault modes of the system. By comparing the model-predicted behavior with the actual observations, a set of consistent diagnosis candidates can be computed. The research focuses on identifying the minimal diagnosis set from all candidates.

Next, the authors introduced the concepts of “sub-observation” and “key sub-observation.” A sub-observation is an abstract representation of the original observation, containing partial information from the original observation. A “sufficient sub-observation” refers to one that can derive the same minimal diagnosis set as the original observation. A key sub-observation is the most abstracted sufficient sub-observation among all sufficient sub-observations.

To compute the key sub-observation, the authors designed a general algorithmic framework. The algorithm starts from the sub-observation corresponding to the original observation and gradually increases the level of abstraction, pruning the branches of sub-observations that are insufficient to derive the minimal diagnosis. Once a sufficient sub-observation is found, the algorithm continues searching based on its “sub-observation” until no further abstraction is possible, thus obtaining a key sub-observation. According to the theorem, the algorithm can provide a correct and complete key sub-observation, and it is guaranteed to terminate.

The researchers further extended the framework to handle diagnosis problems based on state observations and event sequence observations. For state observations, they defined sub-observations as partial value assignments of observation variables; for event observations, sub-observations are subsequences of event sequences. The authors elaborated on how to instantiate the framework in these two cases and introduced the concept of identifying “conflict pairs” to further optimize the search.

This research proposes an effective theoretical framework and algorithm for identifying key diagnostic observational data. This contribution helps provide explainable diagnostic results for artificial intelligence systems and supporting evidence for diagnostic decisions, thereby improving the trustworthiness of AI systems. The framework has a certain level of generality and can be applied to different types of model-based diagnosis problems.