Critical Observations for Model Based Diagnosis: Theory and Practice

Diagnosis in the broadest sense is the determination of the cause and nature of behaviours or observations that deviate from the expected norm. Diagnostic processes identify, categorise, and label these deviations such that they can be addressed and ultimately corrected for. For formalisable systems, the predominant approach to this problem is the application of model-based diagnostics (MBD). Model-based diagnosis takes a simulable facsimile (the model) of a given system and, much like a doctor to a patient, determines the ways the model may have broken down that allow for the observed behaviour of the system to occur. Unlike a doctor however, MBD has not traditionally produced an explanation for the resulting diagnosis. Where a doctor might rely on specific observations (e.g. a high fever) to explain their diagnosis, computational diagnosis uses all available information to draw conclusions. With the increasing complexity of automated systems, and in turn the increasing amounts of generated data, the ability to determine an explanation for a computed diagnosis is arguably now out of reach of the humans whose job it is to take action when deviating behaviour is detected. This thesis addresses the problem of explaining computed diagnoses by introducing Critical Observations. A critical observation is the isolation of only the most relevant observable data once a diagnosis has been determined. We first introduce a mathematical framework and the necessary elements required to automate this process (the theory). Subsequently, we show how this framework can be applied in the two of the most common MBD paradigms (the practice). These two modelling paradigms are known respectively as state-based and event-based. Due to various factors, not least the difficulty in procurement, systems often have no formal model. In these cases, the diagnostic approaches used are collectively referred to as data-driven. Data-driven diagnostics are based on machine learning principles, such as pattern recognition. These approaches typically complement MBD, and we further show how the theory of critical observations can be applied in these settings. In the case of event driven systems, we can build partial discrete event models from the data. These partial models can then be used to improve pattern recognition techniques when the theory of critical observations is applied