Automated qualitative abstraction and its application to automotive systems

The increasing complexity of engineered devices, e.g. in the domain of automotive systems, has lead to an increased demand for computer-supported behavior prediction, diagnosis, and testing. Model-based reasoning is an emerging subfield of Artificial Intelligence that is concerned with representing knowledge about the structure and behavior of physical systems in terms of a model and using it to automate the above-mentioned tasks. Modeling is the hard part of model-based reasoning. In order to make it feasible, it is crucial to break down the knowledge about a device into re-usable components and to organize them in a library. On the other hand, in order to keep reasoning with a model computationally tractable, it is important to have an adequate representation that includes only the distinctions that are required to perform a particular task. In this thesis, we deal with the problem of finding a level of granularity for a behavior model that is as coarse as possible, but still fine enough for a given behavior prediction or diagnosis task. The focus is on task-dependent domain abstraction: i.e., the problem is to determine distinctions within the domains of variables (termed qualitative values) that are both necessary and sufficient, given the constraints of the behavior model, a granularity of possible observations, and a granularity of desired results. We present a method that allows to automatically determine qualitative values, starting from a base model that has been composed from a library. A principled application of this work is to turn real-valued models, as commonly used in industry, into qualitative models to make them accessible to automated reasoning methods. The resulting methods and software tools thus greatly enhance the ability to use a behavior model of an engineered device as a common basis to support different tasks along its life cycle. The thesis describes the application to real-world examples taken from the automotive domain. This leads to the first model-based diagnosis system running on-board a passenger vehicle. The prototype is shown to provide useful results for a number of emission-related failure scenarios that were implemented on a Volvo demonstrator car.