A Research On The Algorithms Of Target Identification Fusion With Bayesian Networks

Target identification fusion based on satellite reconnaissance has become an essential means of getting the controlling power of spacial information. Electronic reconnaissance and infrared photographic reconnaissance are its important and widely-used instruments. There has strong redundancy and complementarity between them in location and identification precision, etc. Intelligentized integration of them could ameliorate the accuracy and reliability of identification and the robustness of the identification system. For passive electronic reconnaissance and infrared photographic reconnaissance, this thesis studies the identification of electromagnetic emitters and platform both by single sensor and by fusion of multisensor data.(1) The uncertainty of the battlefield environment is analyzed. The Bayesian Network Classifiers is introduced to emitters and target identification with single sensor. Several augmented methods to Naive Bayes are compared to found a classifier best fit for the backdrop of target identification with single sensor. The conditional tables (CPTs) are estimated using the Laplace method. The tie-in between targets and equipments were injected into the network by means of CPTs, to conversion from equipment identification into target identification.(2) For dealing with the complexity and uncertainty in the target identification fusion, the fusion model of Bayesian Networks is used. Because of the intractability of construction of unrestricted Bayesian Networks Naive Bayes and angmented Naive Bayes is used as the fusion structure of the fusion network. Expert knowledge and induction from case is combined and TOP-DOWN method is used to construct the dynamic multilayer identification net.(3) According to the analysis of Naive Bayes and angmented Naive Bayes, the Classification-based Super Parent augmenting method is improved.(4) To induce the structure of the net from uncompleted dataset, Gibbs sampling method is used to complete the dataset first. The upper limit and the lower limit is computed first when inducing the CPTs from uncompleted dataset, and the point value is estimated by some method such as averaging.On the backdrop of electronic reconnaissance and infrared photographic reconnaissance, we develop a multisensor target identification system based on Bayesian Networks. And we also compare the result with that based on D-S evidential theory.