Automatic Model Generation for Stochastic Qualitative Reasoning of Building Air Conditioning Systems

In the stochastic qualitative reasoning, which we have proposed, the probabilistic process is used for state transitions based on the stochastic qualitative model . States with relatively small existence probabilities are eliminated in order to suppress the number of generated states under computable order . The determination of stochastic parameters is the most difficult task in model construction, however, because this model many parameters and therefore the information needed to construct it cannot be obtained . This paper proposes a automatic model generation method in order to solve this problem . First, propagation rules and functions are formalized with a few characteristic parameters . As a result, the variable elements of the model can be reduced to less than five percent . Next, a reasonable qualitative model will be generated with measured field data by using the characteristic parameter tuning method . This method will be applied to an actual air conditioning system in a building . A desired qualitative model can be generated in 2.5 lours, it took 8 hours when using the usual method . In addition, institute parameters can be reduced to 25 from 3905. Introduction Qualitative reasoning is a key technology for model based fault detection, in which a section in failure can be identified by comparing the results of reasoning with the real measured values(Kuipers & Berleant 1992)(Lackinger & : Nejdl 1993)(Lackinger & Obreja 1991) . However, the possible behavior patterns of the model tend to increase enormously because of the ambiguity of the qualitative model . In order to solve this problem we have proposed the stochastic qualitative reasoning(Mihara et al . 1994)(Arimoto et al . 1995) . In this method, the probabilistic process is used for state transitions based on the stochastic qualitative model, and states with relatively small existence probabilities are eliminated in order to suppress the number of generated states under computable order . The effectiveness of this method has been shown through simulation experiments for building air conditioning systems(Yumoto at el . 1996a) . However, the model generation of a target system is one of the most difficult problem for qualitative reasoning, because a model has many stochastic parameters and, therefore, the information needed to determine these parameters from the target system instrumentation diagram cannot be obtained . In addition, generated models can be only validated by human intuition . This paper proposes an approach to automatic model generation in order to solve this problem(Yumoto at el . 1996b) . First, propagation rules and functions in a model were formalized with several characteristic parameters from the perspective of regular relation among stochastic parameters . Next, a sensitive analysis for the characteristic parameters was performed on an arc and a function for parameter tuning . Finally, we will propose a scheme for automatic model generation as follows : (1) construct a template of the qualitative model from the target system instrumentation diagram, (2) establish characteristic parameters for the propagation rules on the arc and the functions in the model based on the qualitative knowledge that we have inputted, (3) adjust and determine the values of these parameters based on the measured data by using a steepest ascent based method . This method was applied to a real air conditioning system in a building . We demonstrated the effectiveness of the automatic model generation of a stochastic qualitative model which expresses the normal condition of the target system . real supplied _ air temp .