Research On One Class Of Dynamic Bayesian Network Models And Their Applications

In our daily life and science field, more and more data can be obtained easily.Among all types of data, time series data has been widely used in many applied fields,such as astronomy, geography, biology, physics, chemistry, image processing, speechcommunication, sonar technology, remote sensing technology, nuclear engineering,environmental engineering, medical engineering, ocean engineering, metallurgicalengineering, mechanical engineering, national economy, market economy, productionmanagement, population, and so on.To analyze time series data, many models have been proposed. A dynamicBayesian network model (DBNM) attracts increasing attentions by more and moreresearchers. This thesis mainly studies one class of DBNM. The typical representativesin this class of DBNM, which have both latent random process and observation randomprocess, are hidden Markov model (HMM) and state space model (SSM). Therefore,problems of parameter estimation with HMM and SSM are investigated in this thesis,such as choosing the hidden state number of HMM, state estimation of SSM undernon-Gaussian observation noise, and so on. New methods and algorithms are proposedto deal with the focused problems. Furthermore, some issues in Multi-sensor fusiontracking are also studied.The main contributions of this thesis are given as follows:Firstly, the main research problems of HMM are introduced, especially on theproblem of model selection, which means choosing the hidden state number of HMM.Furthermore, the relationship between feature selection and model selection of HMM isanalyzed. A method of joint feature and model selection of HMM is also proposed. Analgorithm is applied to estimate the parameters of the method. Experiments based onsynthetic and real data demonstrate the effectiveness of the proposed method.Secondly, existing algorithms on state estimation of SSM are studied, especiallyunder non-Gaussian observation noise. The traditional algorithm to estimate the state ofSSM under non-Gaussian observation noise is given as follows: the non-Gaussianobservation noise is modeled by a Gaussian mixture model (GMM), and the state ofSSM is then estimated by expectation maximization (EM) algorithm.However, due to the lack of an effective way to select number of Gaussian in theGMM by the EM algorithm, a variational Bayesian (VB) algorithm is proposed to estimate the state of SSM under non-Gaussian noise in this thesis. Computersimulations show that the proposed approach has an improved performance and a lowercomputation cost compared with the EM algorithm.Thirdly, methods of Multi-sensor single target fusion tracking under non-Gaussiannoise are analyzed. However, using a traditional GMM to model non-Gaussian noiseleads to so many models. It will yield a bad result and high computational time.Therefore, a Student-t distribution is proposed to model non-Gaussian noise inMulti-sensor single target fusion tracking applications. Furthermore, both batch andrecursive algorithms are also provided. It is shown by simulation that the proposedmethods have a more robust estimation performance than the competing methods.Finally, in Multi-sensor Multi-target fusion tracking applications, methods ofsensor relative registration, data association, and sensor fusion are given, respectively.These three processes are dependent on each other. Therefore, a joint sensor relativeregistration, data association, and sensor fusion approach is proposed. The problem ofsensor relative registration, data association, and sensor fusion are then consolidatedinto an estimation problem under a joint optimization framework. The correspondingalgorithm is proposed to estimate the parameters of this optimization framework.Simulation results demonstrate that the proposed method is efficient and is capable ofproviding accurate estimates.