Automatic Parameter Tuning of Stochastic Qualitative Model with Measured Data

Abstract

Stochastic qualitative reasoning is one of key technologies for the model based fault detection, in which a part in failure is identified by comparing the results of reasoning with the real measured values. However, since a model must be constructed with many stochastic parameters, the parameter tuning process is one of the most difficult problem in model generation. In addition, generated models can be only validated by human intuition.
This paper proposes an approach to automatic parameter tuning in order to solve this problem. First, propagation rules and functions in a model are formalized with several characteristic parameters. Next, we have proposed a method for automatic parameter tuning by the steepest ascent based method. In automatic parameter tuning, because most of the processing time is spent calculating agreement rate, the idea to decrease reasoning times are proposed. Finally, this method is applied to a real air conditioning system in a building. As a results, we verify the model with the averaged parameters satisfies the behaviors for a target system for given terms.