DYNAMIC ACROSS-TIME MEASUREMENT INTERPRETATION

Incrementally maintaining a qualitative understanding of physical system behavior based on observations is crucial to tasks such as real-time control, monitoring, and diagnosis. This paper describes the DATMI theory for interpretation tasks. The key idea of DATMI is to dynamically maintain a concise representation of the space of local and global interpretations across time that are consistent with the observations. This representation has two key advantages. First, a set of possible interpretations is more useful than a single (best) candidate for many tasks, such as conservative monitoring. Second, this representation simplifies switching to alternative interpretations when data are faulty or incomplete. Domain-specific knowledge about state and transition probabilities can be used to suggest the interpretation which is most likely. Domain-specific knowledge about durations of states and paths of states can also be used to further constrain the interpretation space. When no consistent interpretation exists, faulty-data hypotheses are generated and then tested by adjusting the interpretation space. The DATMI theory has been tested via implementation and we describe its performance on two examples.