Study of a distributed fusion model and its application in oil distribution forecast

In this paper, a distributed fusion model (DFM) is put forward and a corresponding algorithm is designed to solve the problem of oil distribution forecast. DFM comprises a global fusion center (GFC) and several local fusion units (LFU), which connect with each other through a network. LFU executes fusion computation through two fusion-levels: the feature-level fusion that analyzes qualitative data through a classifying analysis method and extracts quantitative data through a BP neural network method; and the decision-level fusion that conducts decision-level analysis on the results of feature-level fusion through a Bayesian network. DFM decreases global complexity and increases the veracity of the whole system after it increased veracities of local fusion units. The method has been successfully proved in an application to be able to meet the requirement of oil distribution forecast, because it decreases by 47 times more the training cycle than the traditional method -single Bayesian network fusion method - and yielded a higher accuracy rate of oil forecasts than the single neural network fusion method.