Distributed Self-Localization in Self-Organizing Networks

In many commercial and public service scenarios, location awareness is an essential requirement whose importance will grow in the future. The accurate and reliable localization and tracking of mobile network nodes (dynamic self- localization) may be the ultimate goal or a prerequisite for a meaningful interpretation of some collected data. Self- localization methods need to perform well in challenging conditions such as decentralized network architectures, limited internode communication, and node mobility. The elaborate centralized control, significant communications overhead, and lack of scalability of most current wireless networks are avoided by the novel paradigm of self- organization, where only local communication is allowed and each node adapts its procedures based on its own observations and information received from neighboring nodes. Following this paradigm, the goal of the proposed postdoc stay is to devise cooperative and fully distributed algorithms for dynamic self-localization of sensor networks. Instead of using a central processing or control unit, these algorithms rely on an information exchange between neighboring nodes, even if these nodes do not have a precise knowledge of their own location. Computational complexity and communication requirements are to be kept as small as possible. State-of-the-art methods for self-localization are typically not distributed or not dynamic or too demanding with respect to computation and communication. In the proposed project, two different novel approaches are to be pursued and compared. The first approach is based on the Markov chain Monte Carlo concept, more specifically Gibbs sampling and partially collapsed Gibbs sampling. The second approach is based on the distributed total least squares method. Further aspects to be investigated include decentralized scheduling, joint distributed synchronization and localization, the choice of the ranging method, and a performance comparison with centralized algorithms. The proposed research is to be carried out within the Circuits and Systems group at the Delft University of Technology, where considerable expert knowledge on various aspects of self-organizing networks and cooperative self-localization is available.

The main result of this project is the design of a joint concept and method for two tasks: optimal data censoring and rejection of outlier data. These two tasks are highly relevant, for example, for the application of localization in large sensor networks, but also for numerous other applications involving big data. Data censoring is often an essential prerequisite for being able to extract useful information from excessive amounts of data or measurements. Based on some suitable decision criterion, data censoring reduces the available data to a smaller subset of informative data that is used for subsequent analysis (e.g., estimating position coordinates), while the less informative rest of the data is discarded. The design of the decision criterion is critical for the performance that can be achieved in subsequent data processing. The criterion is typically based on statistical prior assumptions about the processes that generate the data, i.e., a stochastic data model. In many cases, however, part of the data do not con form to any prior assumptions but are severely corrupted. Identifying and excluding such corrupted data amounts to our second task, outlier rejection. Existing methods for data censoring are known to be sensitive to outliers, whereas existing methods for outlier rejection are not designed to significantly reduce the data in an optimal way. In the course of the present project, we developed a joint concept that unifies the two tasks to achieve robust censoring. We use the likelihood derived from the assumed uncorrupted data model in order to determine how informative a subset of the available data is, while concurrently excluding outlier data. Differently from many existing methods for outlier rejection, our approach is not confined to particular models of outlier perturbations. We formulate of the problem of robust censoring in terms of non-convex optimization. To solve it, we proposed a Metropolis-Hastings (MH) sampler method that operates on small subsets of the data, thus limiting the computational complexity. MH sampling is a versatile iterative method capable of solving a wide range of problems that are too challenging for most classical estimation methods. The flexibility of the Metropolis-Hastings concept requires that its implementation be carefully designed to ensure convergence within a moderate number of iterations. We developed such a design for the robust censoring problem. Our approach is general in that it is not restricted to any specific data model and does not rely on linearity, uncorrelated measurements, or additive Gaussian noise. Besides this main result, we also achieved advances in three other fields. First, we considered the problem of estimating the parameters of a stationary autoregressive (AR) process when the observed signal is not the AR signal itself but a compressed version of it. We developed a novel approach that uses MH within Gibbs sampling. Second, for the application of cognitive radio, we developed an innovative cooperative spectrum sensing strategy that outperforms existing approaches in terms of both throughput and energy efficiency. This is achieved through a novel optimality criterion with additional degrees of freedom that are neglected in previous designs. Third, we developed a method for Bayesian blind deconvolution of an unknown sparse sequence convolved with an unknown pulse, where the sparse sequence is given an additional structure by enforcing a prescribed minimum distance between the pulse centers.