5. Conclusions
This paper has proposed an intelligent fault diagnosis network for VRF systems using bayesian belief network. Because of its special attribute, the BBN accounts for relationships between variables with probability and distributions. Thus the most important and significant thing for BBN is to obtain the suitable network structure and probability distributions. The subject in this study is VRF system which has seldom been studied. As for its BBN structure, manual setting based on experience and experts’ knowledge and data mining combining feature selection and machine learning are united, which is different from previous researches. On the other hand, the probability distributions are obtained mainly by statistical analysis. Prior probability distributions are often on frequencies in training data set, while the conditional probability tables are calculated by algorithms under some independence assumptions. Besides, the constructed network has been evaluated by test data set. The diagnosis results has evaluated using the part of experiment data, which demonstrates that such a BBN is an excellent tool for fault diagnosis. It is noticeable that it is difficult to separate refrigerant overcharge from normal condition especially when the fault level is not heavy. It is because of the attribute of VRF system. Because there is an accumulator in the system which can accumulate extra refrigerant so that it does not affect performance. But when the surplus exceeds the capacity of accumulator,the fault can be diagnosed. Furthermore, during bayesian inference, the order of input evidence have an impact on fault isolation efficiency in spite of the final posterior probabilities are same.
