From Models to Real-Time: The KSE Nuclear Plant Alarm Processing System

This paper essentially presents an application, namely a real-time model-based monitoring system for nuclear plants. This system is currently being used in two plants in France. We describe how we used various AI techniques for building it: a model-based approach, a logical model of its operation, a declarative implementation of these models, and original knowledge-compiling techniques for automatically generating the real-time expert system from those models. Some of those techniques have just been borrowed from the litterature, but we had to modify or even invent other techniques which simply did not exist. We were pushed forward in doing so by the need of realizing an actual operational system. So, we argue that, considering state-of-the-art AI litterature, when seriously realizing a relatively ambitious application, a large amount of work must be devoted to 'theory' -though of modest scope-, indeed both to theorization of particular aspects, and to integration of various approaches. We think that this reveals a need, which does not seem to be fulfilled today, to redirect the 'theory' of 'applicative AI'.