Qualitative Abstractions for Diagnostics of Intelligent Buildings
Abstract
Although automation and control of technical devices in so called "intelligent buildings" has been rapidly developed in recent years, successfully integrated diagnostics has not yet been produced. Rather, building diagnostics fails in meeting the demands of consistency, steadiness and maintainability. We propose to base diagnostics upon qualitative models that can inexpensively be generated from existing "global" source models. Since however model-based diagnostics (MBD) is in general NP-hard and the size and complexity of building systems differ by orders of magnitude, abstraction becomes an inevitable precondition.
Abstraction cannot be generalized, and an "adequate" method depends on characteristics of the original system model and on the preferences related to the target application. This thesis investigates for classes of building system the effect of abstraction methods on the quality of diagnostics. In order to compare diagnostics models of different paradigms and level of abstraction we introduce a metric for diagnostics quality that is a pair of a measure for complexity and a measure that describes diagnostics fidelity (i.e., fault coverage and diagnosability). This metric is essential for our trade-off analysis: For classes of building systems and diagnostics preferences we explore which approximations produce "proper" models with least possible complexity at required fidelity.
Comprehensively, our intention is a tool that supports a system engineer to generate diagnostics models that balance complexity and fidelity for a given task. We present an implemented transformation from models composed from a global library to models for diagnostics that are based on propositional logic. Components of this global library are described as hybrid automata. To be applicable to diagnostics reasoning, the dynamics of continuously evolving variables must be approximated through a discrete state-space. A sequence of obligatory transformation steps towards a model for diagnostics is developed providing two alternative discretization techniques: sign-based \emph{deviation} (direction of change) and \emph{absolute} (interval) approximation. We further associate one optional model reduction technique to each discretization methods: \emph{energy-based abstraction} to be conducted prior to direction-of-change approximation, and \emph{domain abstraction} after interval approximation. Hence, we can generate highly to sparsely intertwined deviation models and interval-based models of fine to coarse granularity.
Because of the generally lower complexity, we prefer sign-based deviation models to interval-based models wherever applicable, and our results show that deviation models are sufficient for diagnosing most building systems. However, the assumptions are stringent and need to be carefully checked in a real system. Especially diagnostics of systems with only uncontrollable or unobservable continuous dynamics would be more reliable if models describe the absolute values, even with coarse granularity. Our case studies consist of an autonomous lighting system, two lighting and blinding systems of different dimensions and a simple heating system. The experimental results underline our analytical outcome.