Qualitative and Quantitative Probabilistic Temporal Reasoning - for Industrial Applications

Many real-world domains, such as industrial diagnosis, require an adequate representation that combines uncertainty and time. Research in this field involves the development of new knowledge representation and inference mechanisms to deal with uncertainty and time. Current temporal probabilistic models become too complex when used for real world applications. In this paper, we propose a model, Temporal Events Bayesian Networks (TEBN), based on a natural extension of a simple Bayesian network. TEBN tries to make a balance between expressiveness and computational efficiency. Based on a temporal node definition, causal-temporal dependencies are represented by qualitative and quantitative relations, using different time intervals within each variable (multiple granularity). Qualitative knowledge about temporal relations between variables is used to facilitate the acquisition of the quantitative parameters. The inference mechanism combines qualitative and quantitative reasoning. The proposed approach is applied to a thermal power plant through a detailed case study, with promising results.