Decentralized Algorithms Of Cooperative Navigation For Mobile Platforms

Mobile multi-platform systems have been applied widely in both military andcivilian fields. Since navigation is a fundamental capability for a mobile platformand navigation cooperatively of multi-platform system has advantages over inde-pendent navigation of each platform, cooperative navigation is gradually becominga hot research topic. The decentralized data fusion structure, where each platformacts as a processing node and none is central, is a natural choice for cooperativenavigation algorithms of a multi-platform system where the sensor data are dis-tributed among the platforms. Compared with the centralized or hierarchy datafusion structure which relies on a single central node, the decentralized data fusionstructure can enhance the survivability of the system, which has special meaningto military applications. Since the decentralized algorithms require high communi-cation capability, the earlier decentralized cooperative navigation algorithms oftenreduce the communication volume at the price of accuracy. With the developmentof mobile communication technology, it is of great theoretical and practical valuesto design decentralized cooperative navigation algorithms which are as accurate aspossible.The thesis investigates the decentralized cooperative navigation algorithms tomobile multi-platform at two levels. Algorithms are first designed and proposedwith centralized data fusion structure. The algorithms are expected to have localcomputation property to be realized in the distributed computation environment.Decentralized algorithms are then designed and analyzed for the application of mo-bile multi-platform navigation. The essence of the design of the decentralized algo-rithms is to explore the locality of the algorithms and distribute the computationamong the platforms with proper communication strategies.In terms of the design of centralized algorithms, the thesis formulates the statespace model and the probabilistic graphical model, including Markov random field(MRF) and dynamic Bayesian networks (DBNs), of the cooperative navigation prob-lem. The augmented information filter (AIF) is investigated based on the state spacemodel and aided by the MRF. Founded on the DBN, a moments parameterized lazy propagation algorithm for Gaussian Bayesian networks and a novel incremental junc-tion tree algorithm for dynamical inference are proposed. By integrating the twonew algorithms, the inference on Gaussian DBNs can be implemented. Both theAIF and the inference algorithm for Gaussian DBNs achieve the minimum varianceestimate for linear dynamic systems and involve only the linearization errors fornonlinear systems.1. The AIF is designed to solve the cooperative navigation problem. The AIFis derived and expressed with the aid of the MRF. The local computation propertyof the AIF is analyzed. The problem of state recovery is addressed utilizing theCholesky factorization of the information matrix.2. A lazy propagation algorithm with moments parametrization for GaussianBayesian networks is proposed. Di?erent from the conventional junction tree al-gorithms, the lazy algorithm keeps the factorized form of the potential functionsand the messages so as to avoid the complex combination of conditional distribu-tions. The junction tree algorithms with moments parametrization is applicable ina wider range and more numerically stable than that with information parametriza-tion. The design of the novel lazy propagation algorithm provides immediate andcomplete solutions to the inference on Gaussian Bayesian networks.3. A novel incremental junction tree algorithm for dynamical inference is raised.Discrete, continuous and mixed DBNs can all be decomposed into a series of chaingraphs. Passing messages between the junction trees of the consecutive chain graphs,the inference of DBNs can be done incrementally. The junction tree can be rear-ranged by introducing the Push operations, resulting an increase of the freedom ofthe junction tree construction. The conventional forward interface algorithm can beregarded as an application of the novel algorithm with certain constraints. The newalgorithm provides more inference choices for DBNs.In terms of the design and analysis of the decentralized algorithms, the decen-tralized augmented information filter (DAIF) and the decentralized junction tree(DJT) algorithm are proposed. The two decentralized algorithms can obtain thesame accuracy with the corresponding centralized algorithms. A thorough analysisand comparison of the two algorithms and a conventional decentralized solution ispresented. 4. A decentralized augmented information filter to cooperative navigation isdesigned. Each platform fuses the data of local sensors and obtains part of in-formation parameters of the joint state. The platforms implement cooperativelythe distributed incremental Cholesky modification of the information matrix andfurther achieve the estimates of the state moments for each platform. A chain com-munication topology is required in the solution and only two platforms are involvedin the main communication task at a moment.5. A framework of decentralized junction tree algorithm to cooperative navi-gation is proposed. A connection tree is constructed to re?ect the communicationrequirements of the platforms by mapping the cliques of a junction tree to the plat-forms. Each platform initializes the cliques assigned to it with the local sensordata. The platforms accomplish the junction tree algorithm by passing messagesinstructed by the junction tree and the connection tree. The communication topol-ogy is a tree and the communication management is based on the message passingstrategies of the junction tree algorithms.6. A set of performance indicators, composed of accuracy, computation com-plexity, communication complexity and work load balance, special for the decentral-ized algorithms is raised to evaluate various algorithms. The two solutions proposedin the thesis and the conventional decentralized Kalman filter (DKF) solution areanalyzed and compared from various aspects. Compared with the DKF solution, theproposed solutions have much lower requirements of computation synchronizationamong platforms and thus are more feasible to practical applications. Comparedwith the DAIF solution, the DJT solution possesses more distribution structuresand a variety of work load balance to satisfy di?erent requirements of applications.